Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Bayesian Approach to OOD Robustness in Image Classification
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 Unsupervised Generative Transition (UGT), a Bayesian compositional model that learns a transitional vMF kernel dictionary shared between source and unlabeled target domains, plus spatial coefficients from source labels and pseudo-label finetuning, for OOD-robust and occlusion-robust image classification.
Problem
Image classifiers degrade under real-world out-of-domain nuisances (weather, texture, pose, shape, context) and partial occlusion when only unlabeled target-domain images are available.
Approach
Extends CompNets by decoupling unsupervised learning of vMF kernels from supervised learning of spatial coefficients, learning a transitional vMF dictionary regularized toward source kernels so that spatial geometry learned on source transfers to target, then refining with pseudo-labels.
Key results
- On OOD-CV combined nuisances at 0% occlusion, UGT reaches 0.850 top-1 vs 0.840 for the best listed UDA baseline (CST) and 0.720 for CompNet.
- On Occluded-OOD-CV combined nuisances at 60-80% occlusion, UGT achieves 0.501 vs 0.477 for CST.
- On UDAParts to Pascal3D+ at 0% occlusion, UGT reaches 0.992 vs 0.950 for BNA and 0.940 for CompNet (ResNet-50).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UGT instead of CST | OOD-CV Combined · OOD-CV Combined, L0 | top-1 classification accuracy: .840 -> .850 | Helped |
| use UGT instead of the strongest listed baseline | OOD-CV Context · OOD-CV Context, L0 | top-1 classification accuracy: .867 -> .875 | Helped |
| use UGT instead of the strongest listed baseline | OOD-CV Weather · OOD-CV Weather, L0 | top-1 classification accuracy: .813 -> .856 | Helped |
| use UGT instead of the strongest listed baseline | OOD-CV Texture · OOD-CV Texture, L0 | top-1 classification accuracy: .918 -> .936 | Helped |
| use UGT instead of the strongest listed baseline | OOD-CV Pose · OOD-CV Pose, L0 | top-1 classification accuracy: .887 -> .892 | Helped |
| use UGT instead of the strongest listed baseline | OOD-CV Shape · OOD-CV Shape, L0 | top-1 classification accuracy: .838 -> .852 | Helped |
| add the transitional vMF dictionary and spatial coefficients (ablation) | UGT · OOD-CV Combined, L0 | classification accuracy: .698 -> .816 | Helped |
| finetune the spatial coefficients using pseudo-labelled target features (ablation) | UGT with transitional dictionary and spatial coefficients · OOD-CV Combined, L0 | classification accuracy: .816 -> .852 | Helped |
| add the transitional vMF dictionary and spatial coefficients (ablation) | UGT · OOD-CV Texture, L0 | classification accuracy: .715 -> .785 | Helped |
| finetune the spatial coefficients using pseudo-labelled target features (ablation) | UGT with transitional dictionary and spatial coefficients · OOD-CV Texture, L0 | classification accuracy: .785 -> .843 | Helped |
| add the transitional vMF dictionary and spatial coefficients (ablation) | UGT · ImageNet-C Snow, L0 | classification accuracy: .529 -> .781 | Helped |
| finetune the spatial coefficients using pseudo-labelled target features (ablation) | UGT with transitional dictionary and spatial coefficients · ImageNet-C Snow, L0 | classification accuracy: .781 -> .885 | Helped |
Surprising findings
- Many generic UDA methods perform poorly without ImageNet pretraining, while UGT is trained without pretrained backbone features for fairness.
Limitations
- Occluded-OOD-CV uses superimposed cropped occluders rather than real occlusions.
- Mixed results on some nuisance settings where baselines like MIC are competitive at certain occlusion levels.
- Extension to other tasks such as semantic segmentation is only given as a qualitative proof of concept.
Experimental setup
Datasets: OOD-CV ImageNet-C Pascal3D+ UDAParts MS-COCO (occluder crops)
Benchmarks: OOD-CV Occluded-OOD-CV ImageNet-C (Pascal3D+) UDAParts to Pascal3D+
Metrics: top-1 classification accuracy
Baselines: CompNet RPL BNA CDAN MIC CST DINE MCC
Builds on: CompNets generalized cross entropy EM algorithm
Closest papers by meaning
- In Search of Forgotten Domain Generalization (ICLR 2025 · similarity 0.87)
- RUNA: Object-Level Out-of-Distribution Detection via Regional Uncertainty Alignment of Multimodal Representations (AAAI 2025 · similarity 0.86)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.86)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.86)
- Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization (ICLR 2024 · similarity 0.86)
- AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models (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.