Research map / Label-efficient and robust learning / OOD and anomaly detection
CNS-Bench: Benchmarking Image Classifier Robustness Under Continuous Nuisance Shifts
Olaf Dünkel, Artur Jesslen, Jiahao Xie, Christian Theobalt, Christian Rupprecht, 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 |
|---|---|---|---|
| use textual-inversion IN* distribution instead of plain Stable Diffusion prompts (ablation) | image generation · generated images vs ImageNet | FID to ImageNet: 33.8 -> 27.1 | Improved |
| use IN* instead of plain SD (ablation) | image generation · generated images vs ImageNet | ResNet-50 accuracy: 0.68 -> 0.74 | Improved |
| use proposed OOC filter instead of CLIP-based filter (ablation) | OOC filtering · manually labeled set | filter accuracy: 0.65 -> 0.88 | Improved |
| use proposed OOC filter instead of CLIP-based filter (ablation) | OOC filtering · manually labeled set | FPR: 0.36 -> 0.12 | Improved |
| use proposed OOC filter instead of CLIP-based filter (ablation) | OOC filtering · manually labeled set | TPR: 0.90 -> 0.88 | No clear effect |
| switch from supervised ViT to VMamba | classifier architecture · CNS-Bench | rCE: 0.926 -> 0.574 | Improved |
| use DeiT3 modern recipe instead of ViT-B | classifier · CNS-Bench | rCE: 0.926 -> 0.610 | Improved |
| use DINOv1 self-supervised pre-training instead of supervised IN1k | ViT-B · CNS-Bench | rCE: 0.926 -> 0.636 | Improved |
| pre-train on IN21k instead of IN1k (supervised) | ViT-B · CNS-Bench | rCE: 0.926 -> 0.722 | Improved |
| increase DeiT3 size from small to large | DeiT3 · CNS-Bench | rCE: 0.747 -> 0.574 | Improved |
| increase DeiT3 size from large to huge | DeiT3 · CNS-Bench | rCE: 0.574 -> 0.582 | No clear effect |
| replace discriminative classifier with diffusion classifier | DiT classifier · snow and cartoon shifts | average accuracy drop: 0.05 -> 0.106 | Hurt |
Closest papers by meaning
- OOD-CV-v2 : An Extended Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images (TPAMI 2024 · similarity 0.90)
- OOD-CV: A Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images (ECCV 2022 · similarity 0.89)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.88)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.88)
- Foundation Model-oriented Robustness: Robust Image Model Evaluation with Pretrained Models (ICLR 2024 · similarity 0.87)
- A Bayesian Approach to OOD Robustness in Image Classification (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.
