Research map / Label-efficient and robust learning / Self-supervised and contrastive
ONDA-Pose: Occlusion-Aware Neural Domain Adaptation for Self-Supervised 6D 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 ONDA-Pose, which translates unlabeled real images into a synthetic domain via a CAD-like radiance field and uses pseudo pose labels from a global pose refiner to self-supervise a GDR-Net pose estimator for 6D object pose estimation.
Problem
Self-supervised 6D object pose estimation suffers from a domain gap between CAD-synthetic and real images, and translated synthetic images fail to reflect real occlusions.
Approach
Translate real images into the synthetic domain (rather than synthetic into real) by rendering from a CAD-like radiance field conditioned on SfM camera poses, then pseudo-label real images with a pretrained pose estimator plus a globally-consistent pose refiner that exploits multi-view camera poses.
Key results
- On LineMOD, ONDA-Pose achieves 93.2% mean ADD(-S), outperforming all self-supervised baselines (best prior 91.7% TexPose) and fully supervised GDR-Net and PVNet.
- On Occluded-LineMOD, ONDA-Pose reaches 69.3% mean, versus 66.7% for TexPose and 63.3% for SMOC-Net.
- On HomebrewedDB (3 shared objects), ONDA-Pose achieves 95.8% mean, versus 89.1% for TexPose.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add point matching loss Lpm (ablation) | ONDA-Pose · Occluded-LineMOD | ADD/ADD-S recall (%): 21.3 -> 69.3 | Helped |
| add RGB similarity loss Lrgb (ablation) | ONDA-Pose · Occluded-LineMOD | ADD/ADD-S recall (%): 64.7 -> 69.3 | Helped |
| add mask similarity loss Lmask (ablation) | ONDA-Pose · Occluded-LineMOD | ADD/ADD-S recall (%): 65.3 -> 69.3 | Helped |
| add CAD-like radiance field (CARF) (ablation) | ONDA-Pose · Occluded-LineMOD | ADD/ADD-S recall (%): 60.7 -> 69.3 | Helped |
| add global object pose refiner (ablation) | ONDA-Pose · Occluded-LineMOD | ADD/ADD-S recall (%): 67.5 -> 69.3 | Helped |
| use ONDA-Pose instead of the synthetic-data-only backbone estimator | GDR-Net · Occluded-LineMOD | ADD/ADD-S recall (%): 52.9 -> 69.3 | Helped |
Surprising findings
- Disabling the point matching loss drops Occluded-LineMOD mean accuracy from 69.3% to 21.3%, indicating strong instability without it.
Limitations
- Requires multi-view training images with overlapping regions to compute camera poses via COLMAP/SfM.
- Relies on a pretrained backbone pose estimator and mask estimator trained on synthetic PBR data, plus a CAD model.
- Ablation shows the point matching loss is essential and training is unstable without it.
Experimental setup
Datasets: Synthetic PBR Dataset LineMOD Occluded-LineMOD HomebrewedDB
Benchmarks: LineMOD Occluded-LineMOD HomebrewedDB
Metrics: ADD ADD-S
Baselines: Self6D++ SMOC-Net TexPose Self6D DPODv2 GDR-Net CDPN PVNet CosyPose MHP Bousmalis et al.
Builds on: GDR-Net GRAF NeRF Self6D++ TexPose
Closest papers by meaning
- FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (CVPR 2024 · similarity 0.83)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.82)
- Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation (ICLR 2024 · similarity 0.82)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.82)
- Open-Vocabulary Object 6D Pose Estimation (CVPR 2024 · similarity 0.81)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.81)
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.