Research map / 3D vision / Human pose estimation
KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints
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 KeyPose, a transformer-based self-adaptive keypoint detector with keypoint graph feature aggregation and voxel-based NOCS keypoint prediction for category-level 6D object pose estimation from RGB-D.
Problem
Category-level 6D object pose estimation struggles with large appearance and shape variation across instances; prior keypoint methods use local features only and plain cosine similarity for keypoint selection, and direct NOCS keypoint regression is imprecise.
Approach
Predicts object-adaptive keypoints by cross-attending learnable keypoint queries to fused RGB-point cloud features, scores keypoints via convolution-based similarity with randomly sampled feature groups, builds keypoint-cloud and keypoint-keypoint graph features, and predicts NOCS keypoints by classifying voxel locations along each axis.
Key results
- On REAL275, KeyPose reaches 84.2 IoU50, 80.0 IoU75, and 57.7 5°2cm, 66.0 5°5cm (about 6% gain over prior SOTA at 5°5cm per the text).
- On CAMERA25, KeyPose reaches 94.4 IoU50 and 92.6 IoU75, roughly 2% above prior methods.
- On HouseCat6D, KeyPose reports 87.2 IoU25 and 75.4 IoU50, vs. 83.7 and 66.1 for SecondPose; the paper cites over 20% gains on some critical datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the self-adaptive keypoint detection module (ablation) | KeyPose · REAL275 | IoU50 | Helped |
| Use graph-based feature aggregation (ablation) | KeyPose · REAL275 | IoU50 | Helped |
| Use NOCS voxelization (ablation) | KeyPose · REAL275 | IoU50 | Helped |
| Use 96 instead of 16 keypoints (ablation) | KeyPose · REAL275 | IoU50: 82.5 -> 84.2 | Helped |
| Use 96 instead of 32 keypoints (ablation) | KeyPose · REAL275 | IoU50: 83.1 -> 84.2 | Helped |
| Use 96 instead of 64 keypoints (ablation) | KeyPose · REAL275 | IoU50: 83.6 -> 84.2 | Helped |
| Use 96 instead of 128 keypoints (ablation) | KeyPose · REAL275 | IoU50: 83.3 -> 84.2 | Helped |
| Use 32 instead of 8 random-sampling groups (ablation) | KeyPose · REAL275 | IoU50: 82.0 -> 84.2 | Helped |
| Use 32 instead of 16 random-sampling groups (ablation) | KeyPose · REAL275 | IoU50: 82.5 -> 84.2 | Helped |
| Increase random-sampling groups from 32 to 64 (ablation) | KeyPose · REAL275 | IoU50: 84.2 -> 84.2 | No clear effect |
| Increase random-sampling groups from 32 to 96 (ablation) | KeyPose · REAL275 | IoU50: 84.2 -> 84.2 | No clear effect |
| Include the keypoint separation loss Ld (ablation) | KeyPose · REAL275 | IoU50 | Helped |
Limitations
- Relies on Mask-RCNN instance segmentation as a separate upstream step.
- Paper does not report per-module compute or hardware details.
Experimental setup
Datasets: CAMERA25 REAL275 HouseCat6D
Benchmarks: CAMERA25 REAL275 HouseCat6D
Metrics: IoU50 IoU75 5°2cm 5°5cm 10°2cm 10°5cm IoU25
Baselines: SPD SGPA CR-Net RBP-Pose SAR-Net CatFormer NOCS DualPoseNet GPV-Pose HS-Pose VI-Net AG-Pose
Builds on: NOCS CatFormer DGCNN PointNet++ DINOv2
Closest papers by meaning
- Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (CVPR 2024 · similarity 0.90)
- Meta-Point Learning and Refining for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.88)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.87)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (CVPR 2024 · similarity 0.86)
- Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation (ICLR 2025 · similarity 0.86)
- CatFormer: Category-Level 6D Object Pose Estimation with Transformer (AAAI 2024 · 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.