Research map / 3D vision / Human pose estimation
Joint Learning of Pose Regression and Denoising Diffusion with Score Scaling Sampling for Category-level 6D Pose Estimation
Seunghyun Lee, Tae-Kyun Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train encoder with direct 6D pose regression, then jointly train regression and diffusion heads (ablation) | GenPose-style diffusion pose estimator · REAL275 training time | total training time: 11 days -> 7.5 days | Improved |
| joint learning reaches baseline best accuracy (ablation) | joint learning · REAL275 10deg5cm | training time: 11 days -> 0.5 days to reach baseline best | Improved |
| add joint learning (J) vs diffusion-only training, K=50 without ranker (ablation) | GenPose · REAL275 | mAP 5deg2cm: 49.4 -> 50.8 | Improved |
| add time-dependent score scaling guidance for single-pose sampling (ablation) | GenPose · REAL275 K=1 | mAP 5deg2cm: 13.5 -> 52.2 | Improved |
| combine joint learning and score scaling guidance (ablation) | single-pose sampling · REAL275 K=1 | mAP 5deg2cm: 52.2 -> 52.6 | Improved |
| remove energy-based ranker, halving parameters | GenPose pipeline · REAL275 | parameters (M): 4.4 -> 2.2 | Improved |
| add score scaling guidance with DDIM sampler at 500 steps (ablation) | DDIM sampler · REAL275 K=1 | mAP 5deg2cm: 15.9 -> 52.7 | Improved |
| reduce DDIM sampling steps from 500 to 10 with score scaling (ablation) | DDIM + guidance · REAL275 K=1 | mAP 5deg2cm: 52.7 -> 50.8, 2.70 -> 42.9 FPS | Mixed |
| pre-train encoder with classification or reconstruction instead of pose regression (ablation) | GenPose encoder · NOCS convergence | convergence speed | No clear effect |
| geodesic loss instead of L2 loss for rotation regression pre-training (ablation) | regression head · REAL275 | pre-training performance | Improved |
| constant score scaling schedule instead of exponential (ablation) | score scaling guidance · REAL275 symmetric objects | pose distribution diversity: 5deg2cm 52.7 vs 52.6 (no accuracy gain) | Hurt |
| increase w_max from 1 to 4, then beyond 4 (ablation) | exponential score scaling · REAL275 | pose accuracy | Mixed |
Closest papers by meaning
- 6D-Diff: A Keypoint Diffusion Framework for 6D Object Pose Estimation (CVPR 2024 · similarity 0.90)
- Confronting Ambiguity in 6D Object Pose Estimation via Score-Based Diffusion on SE(3) (CVPR 2024 · similarity 0.90)
- Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation (ICCV 2025 · similarity 0.89)
- Generative Category-level Object Pose Estimation via Diffusion Models (NeurIPS 2023 · similarity 0.89)
- Cameras as Rays: Pose Estimation via Ray Diffusion (ICLR 2024 · similarity 0.87)
- Object Pose Estimation via the Aggregation of Diffusion Features (CVPR 2024 · 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.
