Research map / 3D vision / Human pose estimation
Rethinking Correspondence-based Category-Level 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 SpotPose, a two-stage correspondence-based category-level pose estimator that uses pose-invariant point features, keypoint-wise geometric attention, and learned outlier scores to drop outlier correspondences before Umeyama fitting.
Problem
Category-level 6D object pose and size estimation from RGB-D, where correspondence-based methods struggle with pose-sensitive features and outlier correspondences that hurt pose fitting (RANSAC is slow, deep regressors are not robust).
Approach
Combine a pose-invariant PoseInv-PointNet++ (T-Net alignment, no absolute coordinates) with keypoint-level local and global geometric attention using distance/angle descriptors, and explicitly predict keypoint outlier scores so that inlier correspondences alone are fit with Umeyama, avoiding RANSAC.
Key results
- On REAL275, SpotPose reaches 59.7% at 5°2cm versus 54.7% for AG-Pose and 56.2% for SecondPose.
- On HouseCat6D, SpotPose achieves 23.8% at 5°2cm and 77.0% IoU50, exceeding AG-Pose by 12.3 and 14.5 points.
- Removing outlier-aware prediction drops REAL275 5°2cm from 59.7% to 50.5%; Umeyama with outlier removal runs in 1.97 ms versus 4.94 ms with RANSAC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpotPose instead of AG-Pose | category-level object pose estimation · REAL275 | 5°2cm: 54.7 -> 59.7 | Helped |
| Use SpotPose instead of AG-Pose | category-level object pose estimation · CAMERA25 | 5°2cm: 77.8 -> 80.4 | Helped |
| Use SpotPose instead of AG-Pose | category-level object pose estimation · HouseCat6D | 5°2cm: 11.5 -> 23.8 | Helped |
| Replace pose-sensitive PointNet++ with pose-invariant PoseInv-PointNet++ (ablation) | SpotPose · REAL275 | 5°2cm: 51.4 -> 59.7 | Helped |
| Use distance- and angle-based local and global descriptors instead of no descriptors (ablation) | SpotPose · REAL275 | 5°2cm: 55.8 -> 59.7 | Helped |
| Use pose-sensitive relative position embedding instead of Dis-Ang as the local descriptor (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 56.4 | Hurt |
| Remove feature interaction (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 47.7 | Hurt |
| Replace attention with AvgPool for local feature interaction (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 57.7 | Hurt |
| Replace global attention with AvgPool (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 54.4 | Hurt |
| Set geometric interaction blocks to 0 instead of 6 (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 47.7 | Hurt |
| Increase geometric interaction blocks from 6 to 8 (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 57.9 | Hurt |
| Remove outlier-aware correspondence prediction (ablation) | SpotPose · REAL275 | 5°2cm: 59.7 -> 50.5 | Hurt |
Surprising findings
- Adding RANSAC on top of outlier removal does not improve accuracy and only slows pose fitting.
- Replacing the deep MLP pose estimator with Umeyama plus outlier removal gives better 5°2cm accuracy (59.7 vs 57.1).
Limitations
- Relies on offline Mask R-CNN instance masks, so segmentation quality is not handled by the method itself.
- Still falls below GenPose on CAMERA25 at 5°5cm, and the size is regressed separately rather than via correspondences.
Experimental setup
Datasets: CAMERA25 REAL275 HouseCat6D
Benchmarks: CAMERA25 REAL275 HouseCat6D
Metrics: mAP n°m cm IoU25 IoU50 IoU75
Baselines: NOCS SPD SGPA DPDN IST-Net Query6DoF AG-Pose GenPose HS-Pose SecondPose VI-Net GPV-Pose
Builds on: NOCS PointNet++ AG-Pose SecondPose DINOv2 Umeyama algorithm
Closest papers by meaning
- Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation (ICLR 2025 · similarity 0.89)
- Structure-Aware Correspondence Learning for Relative Pose Estimation (CVPR 2025 · similarity 0.89)
- SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.88)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.88)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.87)
- Universal Features Guided Zero-Shot Category-Level Object Pose Estimation (AAAI 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.