Research map / Detection and segmentation / CNN features
Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object 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 AG-Pose, which adaptively detects sparse keypoints per instance and aggregates local and global geometric features into them for prior-free category-level 6D pose and size estimation.
Problem
Dense correspondence-based category-level 6D pose estimation generalizes poorly to unseen instances with large intra-category shape variation because it ignores explicit local and global geometric information.
Approach
Replace dense point correspondences with a small set of instance-adaptive keypoints, detected by attention-conditioned learnable queries with diversity and object-aware chamfer losses, and enrich them with local (KNN relative positions) and global (inter-keypoint relative positions) geometric features before NOCS regression.
Key results
- On REAL275, AG-Pose reaches 54.7% at 5°2cm, surpassing the prior-based DPDN by 8.7 points and the prior-free Query6DoF by 5.7 points.
- On CAMERA25, AG-Pose achieves 91.3% IoU75 and 77.8% at 5°2cm, outperforming Query6DoF by 1.9 points on IoU50 and 3.2 points on IoU75.
- Replacing IAKD with farthest point sampling drops 5°2cm accuracy on REAL275 from 54.7% to 46.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AG-Pose instead of DPDN | category-level pose estimation · REAL275 | 5°2 cm: 46 -> 54.7 | Helped |
| use AG-Pose instead of Query6DoF | category-level pose estimation · CAMERA25 | IoU75: 88.1 -> 91.3 | Helped |
| replace IAKD with FPS keypoint sampling (ablation) | AG-Pose · REAL275 | 5°2 cm: 54.7 -> 46.2 | Hurt |
| use 16, 32, or 64 instead of 96 keypoints (ablation) | AG-Pose · REAL275 | 5°2 cm: 54.7 -> 47.9; 48.8; 51 | Hurt |
| use 128 instead of 96 keypoints (ablation) | AG-Pose · REAL275 | 5°2 cm: 54.7 -> 52.8 | Hurt |
| replace object-aware chamfer loss Locd with chamfer loss Lucd (ablation) | AG-Pose keypoint learning · REAL275 | 5°2 cm: 54.7 -> 49.8 | Hurt |
| add Locd (ablation) | AG-Pose keypoint learning with Ldiv · REAL275 | 5°2 cm: 46.4 -> 54.7 | Helped |
| add Ldiv (ablation) | AG-Pose keypoint learning with Locd · REAL275 | 5°2 cm: 30 -> 54.7 | Helped |
| add both Ldiv and Locd (ablation) | AG-Pose keypoint learning without keypoint losses · REAL275 | 5°2 cm: 29.3 -> 54.7 | Helped |
| add GAFA (ablation) | AG-Pose · REAL275 | 5°2 cm: 47.1 -> 54.7 | Helped |
| add local feature aggregation (ablation) | AG-Pose without local aggregation · REAL275 | 5°2 cm: 49 -> 54.7 | Helped |
| add global feature aggregation (ablation) | AG-Pose without global aggregation · REAL275 | 5°2 cm: 50.1 -> 54.7 | Helped |
Surprising findings
- Only 16 keypoints already reach 47.9% at 5°2cm on REAL275, comparable to several SOTA methods.
Limitations
- Relies on an off-the-shelf MaskRCNN segmentation and the same masks as prior methods.
- Performance sensitive to hyperparameters such as keypoint count and local neighborhood size K.
Experimental setup
Datasets: CAMERA25 REAL275
Benchmarks: CAMERA25 REAL275
Metrics: 3D IoU50 3D IoU75 5deg2cm 5deg5cm 10deg2cm 10deg5cm
Baselines: NOCS DualPoseNet GPV-Pose IST-Net Query6DoF SPD SGPA SAR-Net RBP-Pose DPDN
Builds on: NOCS DPDN Query6DoF PointNet++ Umeyama algorithm
Closest papers by meaning
- KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints (AAAI 2025 · similarity 0.90)
- Meta-Point Learning and Refining for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.88)
- Structure-Aware Correspondence Learning for Relative Pose Estimation (CVPR 2025 · similarity 0.87)
- ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.87)
- Vision Foundation Model Enables Generalizable Object Pose Estimation (NeurIPS 2024 · similarity 0.87)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.86)
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.