Research map / Detection and segmentation / Open-set and few-shot detection
GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement
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 GeoReF, a category-level object pose refinement network that uses an HS-layer with learnable affine transformations and cross-cloud feature transformation, plus shape-prior features for translation and size error prediction, improving CATRE on REAL275 and CAMERA25.
Problem
Category-level object pose refinement is hard due to large intra-category shape variation and discrepancy between the observed object and the category shape prior.
Approach
Replaces CATRE's PointNet encoder with an HS-layer graph-convolution encoder augmented by learnable affine transformations that align geometric features across shapes, adds a cross-cloud transformation that mixes shape-prior feature transforms into observed-cloud features, and feeds prior features into the translation and size branch.
Key results
- On REAL275 refining SPD, the method reaches IoU75 51.8% and 5°2cm 54.4%, versus CATRE's 43.6% and 45.8%.
- Using only 2% of CAMERA25 training images (~5K), the method reaches IoU75 77.5%, exceeding CATRE trained on the full 275K set (76.1%).
- Refining GPV-Pose raises IoU75 from 23.1% to 49.6%, and refining HS-Pose raises IoU75 from 39.1% to 54.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine SPD initial poses with GeoReF instead of CATRE | REAL275 pose refinement · REAL275 | IoU75: 43.6 -> 51.8 | Helped |
| Refine SPD initial poses with GeoReF instead of CATRE | REAL275 pose refinement · REAL275 | 10°2cm: 61.4 -> 71.9 | Helped |
| Add shape-prior features to translation and size estimation (ablation) | CATRE on REAL275 · REAL275 | 5°2cm: 45.8 -> 48.0; 2.2% improvement | Helped |
| Replace PointNet with HS-Encoder (ablation) | CATRE on REAL275 · REAL275 | IoU75: 43.7% -> 37.3% | Hurt |
| Replace PointNet with 3DGCN-Encoder (ablation) | CATRE on REAL275 · REAL275 | 5°5cm: 53.3% -> 43.4% | Hurt |
| Replace PointNet with HS-layer + LAT (ablation) | CATRE + prior in ST branch · REAL275 | IoU75: 45.8 -> 51.0 | Helped |
| Remove LAT on input points (ablation) | GeoReF on REAL275 · REAL275 | IoU75: 51.8 -> 39.3 | Hurt |
| Remove LAT on geometric features (ablation) | GeoReF on REAL275 · REAL275 | 5°2cm: 54.4 -> 47.4 | Hurt |
| Use one shared LAT instead of an independent rotation LAT (ablation) | GeoReF on REAL275 · REAL275 | 5°2cm: 54.4 -> 50.4 | Hurt |
| Replace cross-cloud transformation with global concatenation fusion (ablation) | GeoReF on REAL275 · REAL275 | 5°2cm: 54.4 -> 47.8; 4.6% decreased | Hurt |
| Use cross-cloud transformation for feature fusion (ablation) | GeoReF on REAL275 · REAL275 | IoU75: 48.4 -> 51.8 | Helped |
| Train with GeoReF instead of CATRE using 5K images | CAMERA25 pose refinement · CAMERA25, 5K training images | IoU75: 63.2 -> 77.5 | Helped |
Surprising findings
- Directly replacing CATRE's PointNet encoder with 3D graph convolution encoders (HS-Encoder or 3DGCN-Encoder) degrades refinement performance.
- CATRE fails to refine Self-DPDN initial poses and degrades HS-Pose estimates at the 10°2cm and 10°5cm metrics.
Limitations
- Evaluated on rigid categories only; extension to articulated objects is proposed but not tested.
- Refinement is iterative, and the paper does not report runtime or efficiency numbers in the main text.
Experimental setup
Datasets: REAL275 CAMERA25
Benchmarks: REAL275 CAMERA25
Metrics: IoU50 IoU75 5°2cm 5°5cm 10°2cm 10°5cm mAP
Baselines: CATRE SPD HS-Pose GPV-Pose Self-DPDN RBP-Pose SGPA DualPoseNet CR-Net NOCS
Builds on: CATRE HS-layer 3DGCN NOCS
Closest papers by meaning
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.89)
- The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement (CVPR 2024 · similarity 0.88)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.88)
- SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.87)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.87)
- Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation (ICLR 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.