Research map / 3D vision / Human pose estimation
Learning Shape-Independent Transformation via Spherical Representations for 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 SpherePose, which projects SO(3)-invariant point and image features onto HEALPix sphere anchors and predicts spherical NOCS correspondences with a transformer and a hyperbolic correspondence loss for category-level pose estimation.
Problem
Category-level object pose estimation via correspondences suffers from shape-dependent canonical (NOCS) coordinates, causing semantic incoherence across object shapes and large NOCS errors.
Approach
Use a sphere (HEALPix-discretized) as a shared proxy shape: project SO(3)-invariant point-wise features onto spherical anchors and predict shape-independent spherical NOCS coordinates via transformer attention, supervised by a hyperbolic (arcosh) correspondence loss.
Key results
- On REAL275, SpherePose reaches 58.2% on 5°2cm and 88.2% on 10°5cm, outperforming AG-Pose (54.7% and 83.1%).
- On HouseCat6D, SpherePose achieves 72.2% IoU50 and 19.3% 5°2cm, surpassing SecondPose (66.1% IoU50) and AG-Pose (62.5% IoU50).
- Mean NOCS angle error drops to 4.52° versus 10.36° for AG-Pose and 16.62° for DPDN on REAL275.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use spherical representations with SpherePose instead of the point-based AG-Pose baseline | REAL275 · REAL275 | mAP (5°2cm): 54.7 -> 58.2 | Helped |
| Use SpherePose rather than AG-Pose | CAMERA25 · CAMERA25 | mAP (IoU 50): 93.8 -> 94.8 | Helped |
| Use SpherePose rather than AG-Pose | HouseCat6D · HouseCat6D | mAP (IoU 50): 62.5 -> 72.2 | Helped |
| Add ColorPointNet++ instead of using no deep point cloud backbone (ablation) | SpherePose · REAL275 | mAP (5°2cm): 55.9 -> 58.2 | Helped |
| Replace PointNet++ with SO(3)-invariant ColorPointNet++ (ablation) | SpherePose · REAL275 | mAP (10°5cm): 84.4 -> 88.2 | Helped |
| Add the Transformer encoder for spherical feature interaction (ablation) | SpherePose · REAL275 | mAP (5°2cm): 20.7 -> 53.6 | Helped |
| Increase Transformer encoder depth from 6 to 8 layers (ablation) | SpherePose · REAL275 | mAP (5°2cm): 58.2 -> 57.9 | Hurt |
| Use hyperbolic L1 instead of L1 correspondence loss (ablation) | SpherePose · REAL275 | mAP (5°2cm): 56.0 -> 58.0 | Helped |
| Use hyperbolic L2 instead of Euclidean L2 correspondence loss (ablation) | SpherePose · REAL275 | mAP (5°2cm): 54.2 -> 58.2 | Helped |
| Replace HEALPix with equirectangular spherical grids (ablation) | SpherePose · REAL275 | mAP (10°5cm): 88.2 -> 83.6 | Hurt |
| Replace HEALPix with Super-Fibonacci Spirals (ablation) | SpherePose · REAL275 | mAP (10°5cm): 88.2 -> 87.2 | Hurt |
| Add low-level RGB image features (ablation) | SpherePose · REAL275 | mAP (5°2cm): 56.8 -> 58.2 | Helped |
Surprising findings
- PointNet++ point-cloud features perform worse than using no point cloud backbone at all on 10°5cm (84.4 vs 86.7), while the SO(3)-invariant ColorPointNet++ improves results.
- HP-PPF, an SO(3)-invariant parameter-free feature, degrades performance relative to no point-cloud backbone.
Limitations
- Transformer attention scales quadratically with the number of spherical anchors, limiting grid resolution (M=768).
- DINOv2 features are only approximately SO(3)-invariant.
- Translation and size are predicted by a separate lightweight PointNet++ rather than the main method.
Experimental setup
Datasets: CAMERA25 REAL275 HouseCat6D
Benchmarks: CAMERA25 REAL275 HouseCat6D
Metrics: mAP at n°m cm (5°2cm, 5°5cm, 10°2cm, 10°5cm) IoU25 IoU50 IoU75 mean NOCS angle and distance error
Baselines: GPV-Pose HS-Pose GenPose DualPoseNet VI-Net SecondPose NOCS SPD SGPA DPDN IST-Net Query6DoF
Builds on: NOCS VI-Net DPDN PointNet++ DINOv2
Closest papers by meaning
- Rethinking Correspondence-based Category-Level Object 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)
- GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement (CVPR 2024 · similarity 0.87)
- Improving Semantic Correspondence with Viewpoint-Guided Spherical Maps (CVPR 2024 · similarity 0.87)
- Learning SO(3)-Invariant Semantic Correspondence via Local Shape Transform (CVPR 2024 · similarity 0.87)
- KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints (AAAI 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.