Research map / 3D vision / Human pose estimation
Meta-Point Learning and Refining 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 MetaPoint, which learns class-agnostic meta-points from learnable embeddings without support, then assigns and refines them to desired keypoints using support keypoints, for category-agnostic pose estimation on MP-100.
Problem
Category-agnostic pose estimation (CAPE) must localize arbitrary keypoints given a few annotated support images, but existing methods rely only on local support features and neglect inherent keypoint information.
Approach
Learnable meta-embeddings interact with query image features via a progressive deformable point decoder to predict support-free potential keypoints (meta-points); these are then assigned to support identities by bipartite matching and refined using support features and meta-embeddings.
Key results
- MetaPoint+ achieves 86.17% average PCK@0.2 in 1-shot on MP-100, versus 85.31% for CapeFormer.
- MetaPoint+ reaches 90.02% average PCK@0.2 in 5-shot, versus 89.30% for CapeFormer.
- Without any support, MetaPoint achieves 79.31% average 1-shot PCK@0.2, comparable to POMNet's 79.70%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use meta-point learning and support-based refinement | CAPE on MP-100 · MP-100 1-shot | PCK@0.2: 85.31 -> 86.17 | Helped |
| Use meta-point learning and support-based refinement | CAPE on MP-100 · MP-100 5-shot | PCK@0.2: 89.30 -> 90.02 | Helped |
| Use meta-point learning and support-based refinement | CAPE · Human Body cross-super-category | PCK@0.2: 83.44 -> 84.32 | Helped |
| Use meta-point learning and support-based refinement | CAPE · Human Face cross-super-category | PCK@0.2: 80.96 -> 82.21 | Helped |
| Use meta-point learning and support-based refinement | CAPE · Vehicle cross-super-category | PCK@0.2: 45.40 -> 46.51 | Helped |
| Use meta-point learning and support-based refinement | CAPE · Furniture cross-super-category | PCK@0.2: 52.49 -> 53.67 | Helped |
| Predict meta-points without support or refinement (ablation) | MetaPoint on MP-100 · MP-100 1-shot | PCK@0.2: 85.31 -> 79.31 | Hurt |
| Add visibility prediction (ablation) | MetaPoint on MP-100 · MP-100 1-shot Split1 | mPCK: 68.18 -> 70.37 | Helped |
| Add support features to refinement (ablation) | MetaPoint on MP-100 · MP-100 1-shot Split1 | mPCK: 70.37 -> 74.70 | Helped |
| Add assigned support meta-embeddings to refinement (ablation) | MetaPoint on MP-100 · MP-100 1-shot Split1 | mPCK: 74.70 -> 76.43 | Helped |
| Add slacked regression loss (ablation) | MetaPoint on MP-100 · MP-100 1-shot Split1 | mPCK: 76.43 -> 77.11 | Helped |
| Use query and support cost matrices together for assignment (ablation) | MetaPoint+ · MP-100 1-shot Split1 | mPCK: 77.11 -> 73.21 | Hurt |
Surprising findings
- Meta-points predicted without any support already achieve performance close to support-based POMNet in 1-shot setting.
Limitations
- Gains over CapeFormer in PCK@0.2 are modest and performance approaches saturation at this threshold.
- Performance varies greatly across super-categories in cross super-category evaluation.
Experimental setup
Datasets: MP-100
Benchmarks: MP-100
Metrics: PCK@0.2 mPCK
Baselines: ProtoNet MAML Fine-tune POMNet CapeFormer
Builds on: CapeFormer POMNet Deformable DETR
Closest papers by meaning
- CapeX: Category-Agnostic Pose Estimation from Textual Point Explanation (ICLR 2025 · similarity 0.92)
- ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.92)
- Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (CVPR 2024 · similarity 0.88)
- KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints (AAAI 2025 · similarity 0.88)
- Recurrent Feature Mining and Keypoint Mixup Padding for Category-Agnostic Pose Estimation (CVPR 2025 · similarity 0.88)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.87)
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.