Research map / 3D vision / Human pose estimation
GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence
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 GigaPose, a CAD-based novel object pose estimator that retrieves templates over only azimuth/elevation and recovers in-plane rotation, scale and 2D translation from a single 2D-2D patch correspondence with RANSAC, achieving ~35x faster coarse pose and better robustness on BOP.
Problem
CAD-based novel object 6D pose estimation suffers from slow template matching (linear in number of templates) and sensitivity to segmentation errors caused by occlusion.
Approach
Use a small set of 162 out-of-plane templates matched via fast nearest-neighbor search in contrastively trained local ViT features (DINOv2-initialized), then estimate in-plane rotation and scale with MLPs from features of a single 2D-2D patch correspondence, and 2D translation directly from patch locations, combined with RANSAC.
Key results
- Coarse pose estimation takes 48 ms per detection versus 1.68 s for MegaPose (about 35x speedup).
- With CNOS detections and GenFlow refinement with 5 hypotheses, mean AR on BOP seven core datasets is 60.5 versus 57.0 for MegaPose.
- Without refinement, mean AR is 26.8 vs 20.8 for MegaPose using CNOS detections.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Fae instead of using DINOv2 features without fine-tuning (ablation) | GigaPose · LM-O, T-LESS, YCB-V mean | AR: 19.0 -> 27.9 | Helped |
| estimate in-plane rotation directly rather than with dense 3DoF templates (ablation) | GigaPose · LM-O, T-LESS, YCB-V mean | AR: 22.1 -> 27.9 | Helped |
| use single-correspondence 2D-to-2D pose estimation instead of 3D-to-2D correspondences with ePnP (ablation) | GigaPose · LM-O, T-LESS, YCB-V mean | AR: 26.5 -> 27.9 | Helped |
| use single-correspondence estimation instead of two 2D-to-2D correspondences with the 2D Kabsch variant (ablation) | GigaPose · LM-O, T-LESS, YCB-V mean | AR: 27.2 -> 27.9 | Helped |
| increase the template count from 162 to 576 (ablation) | GigaPose · LM-O, T-LESS, YCB-V mean | AR: 27.9 -> 28.5 | Helped |
| replace MegaPose with GigaPose for coarse pose estimation | CNOS detections · BOP seven-dataset mean | AR: 20.8 -> 26.8 | Helped |
| replace MegaPose with GigaPose for coarse pose estimation | CNOS detections · YCB-V | AR: 28.1 -> 27.8 | Hurt |
| replace MegaPose with GigaPose before MegaPose refinement | CNOS detections with MegaPose refinement · BOP seven-dataset mean | AR: 50.9 -> 54.7 | Helped |
| replace MegaPose with GigaPose before GenFlow refinement with 5 hypotheses | CNOS detections with GenFlow refinement · BOP seven-dataset mean | AR: 57.0 -> 60.5 | Helped |
| use GigaPose instead of MegaPose with a predicted single-image 3D model | LM-O, predicted 3D model · LM-O | AR: 25.2 -> 27.2 | Helped |
| use GigaPose instead of MegaPose with a predicted single-image 3D model after refinement | LM-O, predicted 3D model · LM-O, refined | AR: 22.7 -> 29.4 | Helped |
| use GigaPose rather than MegaPose for coarse pose estimation | per-detection coarse pose estimation · BOP datasets | runtime: 1.68 s vs 48 ms | Helped |
Surprising findings
- Using 576 templates as in MegaPose improves AR by only 0.6% over 162 templates, suggesting correspondences reduce template memory needs.
- Estimating in-plane rotation with dense templates (5832 per object) decreases AR by 5.8% compared to direct MLP prediction.
Limitations
- Performance degrades on LM-O with small segments, heavy occlusion and low-fidelity CAD models.
- Refinement gains are smaller than for MegaPose, since MegaPose's refiner was trained with coarse-like pose perturbations.
- Coarse pose stability with predicted 3D models depends on careful reference image selection for Wonder3D.
Experimental setup
Datasets: MegaPose synthetic training renderings (BOP) Google Scanned Objects ShapeNet
Benchmarks: LM-O T-LESS TUD-L IC-BIN ITODD HB YCB-V
Metrics: AR (average recall) AP run-time
Baselines: MegaPose ZS6D OSOP
Builds on: MegaPose DINOv2 CNOS InfoNCE
Closest papers by meaning
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.89)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.87)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.87)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.87)
- HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation (CVPR 2024 · similarity 0.86)
- MFOS: Model-Free & One-Shot Object Pose Estimation (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.