Research map / 3D vision / Human pose estimation
Confronting Ambiguity in 6D Object Pose Estimation via Score-Based Diffusion on SE(3)
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.
Applies score-based diffusion directly on the SE(3) group to jointly model rotation and translation for single-RGB 6D object pose estimation under symmetry/occlusion ambiguity, using a closed-form surrogate Stein score.
Problem
Pose ambiguity from object symmetries and occlusions in 6D object pose estimation from single RGB images, where one image maps to many valid poses; existing density methods are confined to SO(3) or are computationally heavy.
Approach
Score-based diffusion on the SE(3) Lie group with a Gaussian perturbation kernel and a surrogate Stein score (-z/σ²) that gives a closed form, avoiding autograd and Jacobian computation, and jointly modeling rotation and translation rather than treating them independently as in R3SO(3).
Key results
- On SYMSOL, Ours (ResNet50) reaches 0.37 average spread in degrees vs. 0.70 for normalizing flows and 1.94 for HyperPosePDF.
- On T-LESS (average of 30 objects), Ours (SE(3)) reaches 93.16 MSPD and 86.94 R@5 accuracy, versus 90.17 MSPD for GDRNPP.
- On SYMSOL-T, Ours (SE(3)) gives rotation spread around 0.4-0.6 degrees across all shapes, outperforming R3SO(3) score model on most shapes, while the R3SO(3) variant is better on translation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint SE(3) diffusion instead of separate R3SO(3) diffusion (ablation) | Score-based pose diffusion · T-LESS | R@2: 27.98 -> 47.21 | Helped |
| Use joint SE(3) diffusion instead of separate R3SO(3) diffusion (ablation) | Score-based pose diffusion · T-LESS | T@2: 60.37 -> 71.72 | Helped |
| Use SE(3) diffusion instead of R3SO(3) diffusion (ablation) | Score-based pose diffusion · SYMSOL-T | rotation spread: 1.38 -> 0.59 | Helped |
| Use SE(3) diffusion instead of R3SO(3) diffusion (ablation) | Score-based pose diffusion · SYMSOL-T | translation spread: 0.017 -> 0.016 | Mixed |
| Use a ResNet50 backbone instead of ResNet34 (ablation) | Ours on SYMSOL · SYMSOL | average angular spread (degrees): 0.42 -> 0.37 | Helped |
| Use the surrogate Stein score instead of the score computed by automatic differentiation (ablation) | SE(3) score-based diffusion · Denoising process | convergence | Helped |
| Use Fourier-based conditioning instead of scale-bias conditioning (ablation) | Score model on SO(3) · SYMSOL | ability to capture periodic score-field features | Helped |
| Reduce SE(3) denoising steps from 100 to 5 (ablation) | Ours (SE(3)) · T-LESS | MSPD: 93.16 -> 92.40 | Hurt |
| Reduce R3SO(3) denoising steps from 100 to 5 (ablation) | Ours (R3SO(3)) · T-LESS | MSPD: 85.73 -> 85.67 | No clear effect |
| Use diffusion-based pose estimation instead of GDRNPP | Ours (SE(3)) · T-LESS | MSPD: 90.17 -> 93.16 | Helped |
| Use diffusion-based pose estimation instead of GDRNPP | Ours (SE(3)) · T-LESS | T@5: 96.09 -> 92.03 | Hurt |
| Use score-based diffusion instead of Normalizing Flows | Ours (ResNet50) · SYMSOL | average angular spread (degrees): 0.70 -> 0.37 | Helped |
Surprising findings
- The R3SO(3) score model is better on translation than the joint SE(3) model, despite SE(3) winning on rotation.
- The SE(3) model's inference runs up to 250 FPS at 5 denoising steps, while the R3SO(3) variant reaches 307 FPS.
Limitations
- Requires ground-truth bounding boxes and visible-part segmentation masks for T-LESS.
- Single model trained across all shapes can degrade performance on some shapes (e.g., cone with ResNet50).
- Translation accuracy is slightly worse than GDRNPP, attributed to GDRNPP's use of 3D model geometry.
Experimental setup
Datasets: SYMSOL SYMSOL-T T-LESS
Benchmarks: SYMSOL SYMSOL-T T-LESS
Metrics: average angular spread in degrees rotation error translation error MSPD MSSD VSD R@2/R@5/R@10 T@2/T@5/T@10 FPS
Baselines: DBN Implicit-PDF HyperPosePDF Normalizing Flows Regression Iterative regression GDRNPP Ours (R3SO(3))
Builds on: score-based generative modeling (NCSN) geodesic random walk Implicit-PDF R3SO(3) diffusion (Urain et al.)
Closest papers by meaning
- 6D-Diff: A Keypoint Diffusion Framework for 6D Object Pose Estimation (CVPR 2024 · similarity 0.92)
- Object Pose Estimation via the Aggregation of Diffusion Features (CVPR 2024 · similarity 0.89)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.87)
- iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.86)
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 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.