Research map / Label-efficient and robust learning / OOD and anomaly detection
OOD-CV: A Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images
Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, Adam Kortylewski
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 |
|---|---|---|---|
| add AugMix training (ablation) | ResNet50 · OOD-CV image classification (texture) | top-1 accuracy: 76.2% -> 82.1% | Improved |
| add AugMix training (ablation) | ResNet50 · OOD-CV image classification (weather) | top-1 accuracy: 69.2% -> 75.2% | Improved |
| add AugMix training (ablation) | ResNet50 · OOD-CV image classification (shape, pose) | top-1 accuracy: 73.2% -> 73.0% (shape), 73.8% -> 73.3% (pose) | Hurt |
| add style transfer augmentation (ablation) | Faster-RCNN · OOD-CV object detection (context) | mAP@50: 35.6% -> 39.4% | Improved |
| add style transfer augmentation (ablation) | Faster-RCNN · OOD-CV object detection (shape) | mAP@50: 61.6% -> 59.8% | Hurt |
| add adversarial training (ablation) | ResNet50 · OOD-CV image classification (i.i.d) | top-1 accuracy: 85.2% -> 83.7% | Hurt |
| replace ResNet-50 backbone with Swin-T | image classification, detection, pose models · OOD-CV all nuisances | accuracy | No clear effect |
| add AugMix to MobileNetv3-L instead of ResNet50 (ablation) | MobileNetv3-L · OOD-CV image classification (context) | top-1 accuracy: 75.9% -> 76.8% (MbNetv3-L) vs 78.7% -> 82.6% (ResNet50) | Mixed |
| use one-stage RetinaNet instead of two-stage Faster-RCNN | detector · OOD-CV object detection (i.i.d) | mAP@50: 72.6% -> 74.7% | Improved |
| use NeMo with explicit 3D geometry instead of Res50-Specific | 3D pose estimation · OOD-CV 3D pose estimation (pose) | Acc@pi/6: 45.2% -> 56.9% | Improved |
| use NeMo instead of Res50-Specific | 3D pose estimation · OOD-CV 3D pose estimation (shape) | Acc@pi/6: 43.5% -> 51.7% | Improved |
| add AugMix to NeMo (ablation) | NeMo · OOD-CV 3D pose estimation (weather) | Acc@pi/6: 49.8% -> 56.7% | Improved |
Closest papers by meaning
- A Bayesian Approach to OOD Robustness in Image Classification (CVPR 2024 · similarity 0.91)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.89)
- Unexplored Faces of Robustness and Out-of-Distribution: Covariate Shifts in Environment and Sensor Domains (CVPR 2024 · similarity 0.89)
- COUNTS: Benchmarking Object Detectors and Multimodal Large Language Models under Distribution Shifts (CVPR 2025 · similarity 0.87)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.87)
- RUNA: Object-Level Out-of-Distribution Detection via Regional Uncertainty Alignment of Multimodal Representations (AAAI 2025 · 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.
