Research map / Label-efficient and robust learning / Domain adaptation and generalization
Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level 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 3DUDA, a source-free, image-only unsupervised domain adaptation method for category-level 3D pose estimation that updates locally robust neural mesh vertex features and the CNN backbone in an EM-style loop.
Problem
Category-level 3D pose estimation degrades under out-of-distribution nuisances (weather, texture, shape, occlusion, corruptions), and existing adaptation methods need target 3D data, depth, or source data; the paper targets source-free, RGB-only unsupervised adaptation.
Approach
Represent categories as cuboid neural meshes with per-vertex vMF features, and adapt by selectively updating only vertex features that remain similar to target image features (locally robust parts), rather than global pseudo-labeling of whole images, alternating with backbone fine-tuning in an EM fashion.
Key results
- On OOD-CV combined nuisance, pi/6 accuracy of 94.0% versus 51.8% for Res50-General and 50.0% for DMNT.
- On Corrupted-Pascal3D+ (Gaussian noise, shot noise, etc.), reported pi/6 accuracy of roughly 84-89% versus NeMo's 43-77%.
- On extreme UDA (OOD+SN), pi/6 accuracy of 80.5% versus NeMo's 32.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt the source model with selective vertex feature updates | NeMo source model · OOD-CV Combined | π/6 accuracy: 51.8 -> 94.0 | Helped |
| adapt the source model with selective vertex feature updates | NeMo source model · Corrupted-Pascal3D+ Gaussian Noise | π/6 accuracy: 43.7 -> 84.3 | Helped |
| adapt the source model with selective vertex feature updates | NeMo source model · Occluded-OOD-CV OccL1 | π/6 accuracy: 30.6 -> 84.6 | Helped |
| adapt the source model with selective vertex feature updates | NeMo source model · OOD-CV + OccL1 + Speckle Noise | π/6 accuracy: 18.6 -> 69.4 | Helped |
| vary clutter model number N across 1, 5, 20, and 35 (ablation) | 3DUDA · OOD-CV bus category | π/6 accuracy: 95.1; 94.4; 95.5; 94.5 | Mixed |
| vary clutter model number N across 1, 5, 20, and 35 (ablation) | 3DUDA · OOD-CV bus category | median error: 3.01; 3.03; 2.99; 2.79 | Mixed |
| increase clutter model number from 5 to 20 (ablation) | 3DUDA · OOD-CV bus category | π/18 accuracy: 91.1 -> 92.3 | Helped |
| increase clutter model number from 20 to 35 (ablation) | 3DUDA · OOD-CV bus category | π/18 accuracy: 92.3 -> 91.7 | Hurt |
Surprising findings
- Generic ResNet-50 classifier performs quite well relative to NeMo and DMNT under OOD-CV nuisances.
Limitations
- Relies on a source model and cannot trivially extend to articulated objects.
- Inference requires multiple pre-rendered samples and render-and-compare optimization, which is costly.
- Concentration parameter kappa is fixed and its importance is not studied in depth.
Experimental setup
Datasets: Pascal3D+ OOD-CV ImageNet-C Occluded-OOD-CV ShapeNet
Benchmarks: OOD-CV Corrupted-Pascal3D+ Occluded-OOD-CV
Metrics: pi/6 accuracy pi/18 accuracy median error
Baselines: NeMo DMNT SyntheticP3D MaskRCNN Res50-General ViT-b-16
Builds on: NeMo DMNT Neural Mesh Models
Closest papers by meaning
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.91)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.89)
- A Dual-Augmentor Framework for Domain Generalization in 3D Human Pose Estimation (CVPR 2024 · similarity 0.87)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.87)
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.87)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 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.