Research map / 3D vision / Human pose estimation
Dynamic Support Information Mining for Category-Agnostic Pose Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SDPNet, a support-based dynamic perception network for category-agnostic pose estimation that mines structural and semantic information from support keypoints to guide query keypoint refinement, outperforming CapeFormer on MP-100.
Problem
Category-agnostic pose estimation (CAPE) predicts keypoints of unseen categories from few annotated support images; existing methods use support samples mainly to initialize query keypoint features and neglect their fine-grained semantic and prior structural information.
Approach
Uses cross-instance self-attention over support keypoints to build category-consensus semantic and structure-aware features; a GCN with a dynamically predicted category-specific prior structure passes information between query keypoints, and support features dynamically modulate high-resolution query feature activation and deformable sampling during iterative refinement.
Key results
- SDPNet (HRNet-32) reaches 87.36% average PCK in 1-shot on MP-100, versus 85.31% for CapeFormer.
- SDPNet (HRNet-32) reaches 90.42% average PCK in 5-shot on MP-100, versus 89.30% for CapeFormer.
- Ablation on split 1 (3-shot): cross-instance interaction, high-resolution dynamic convolution, and prior-structure GCN each add to PCK, reaching 93.54 from 92.97 baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SDPNet support-based dynamic perception | CapeFormer · MP-100 1-shot average | PCK: 85.31 -> 87.36 | Helped |
| Add SDPNet support-based dynamic perception | CapeFormer · MP-100 5-shot average | PCK: 89.30 -> 90.42 | Helped |
| Add weighted support aggregation and cross-instance interaction instead of averaging (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 92.97 -> 93.21 -> 93.54 | Helped |
| Add high-resolution query features (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 92.84 -> 93.38 | Helped |
| Add dynamic convolution to high-resolution query features (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 93.38 -> 93.54 | Helped |
| Add graph convolution with predicted prior structure (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 93.01 -> 93.54 | Helped |
| Use multi-head attention rather than global or concatenation regression to estimate prior structure (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 93.17; 93.48 -> 93.54 | Helped |
| Add the sparse constraint to predicted prior structure (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 92.87 -> 93.54 | Helped |
| Use dynamic modulation matrices generated from keypoint self-information rather than shared or fixed transformations (ablation) | SDPNet graph convolution · MP-100 split 1, 3-shot | PCK: 93.15; 93.22 -> 93.54 | Helped |
| Generate deformable-attention modulation from support-query feature differences rather than support features alone (ablation) | SDPNet · MP-100 split 1, 3-shot | PCK: 93.06 -> 93.32 | Helped |
| Use a shared modulation coefficient for all offsets of a keypoint rather than a unique coefficient per offset (ablation) | SDPNet deformable attention · MP-100 split 1, 3-shot | PCK: 93.32 -> 93.54 | Helped |
Surprising findings
- Generating modulation coefficients from semantic support features alone performed worse than standard deformable attention, which the authors attribute to the lack of query information.
Limitations
- Does not explicitly model the 3D structure of objects, leading to weaker accuracy on some occluded keypoints.
- Predicted prior structures are less reliable for objects with few structural characteristics, such as clothing.
Experimental setup
Datasets: MP-100
Benchmarks: MP-100
Metrics: PCK@0.2
Baselines: ProtoNet MAML Fine-tune POMNet CapeFormer
Builds on: CapeFormer POMNet deformable attention modulated GCN
Closest papers by meaning
- DynPose: Largely Improving the Efficiency of Human Pose Estimation by a Simple Dynamic Framework (CVPR 2025 · similarity 0.83)
- ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.81)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.81)
- Meta-Point Learning and Refining for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.81)
- Recurrent Feature Mining and Keypoint Mixup Padding for Category-Agnostic Pose Estimation (CVPR 2025 · similarity 0.81)
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.