Research map / Label-efficient and robust learning / OOD and anomaly detection
The Invisible Gorilla Effect in Out-of-distribution Detection
Harry Anthony, Ziyun Liang, Hermione Warr, Konstantinos Kamnitsas
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 |
|---|---|---|---|
| make OOD ink artefact dissimilar to lesion ROI colour (vs similar) | Mahalanobis · ISIC ink | AUROC (%): 76.98 -> 63.64 | Hurt |
| make OOD ink artefact dissimilar to lesion ROI colour (vs similar) | KNN · ISIC ink | AUROC (%): 85.68 -> 70.10 | Hurt |
| make OOD ink artefact dissimilar to lesion ROI colour (vs similar) | FeatureNorm · ISIC ink | AUROC (%): 75.12 -> 52.91 | Hurt |
| make OOD ink artefact dissimilar to lesion ROI colour (vs similar) | DDPM-MSE · ISIC ink | AUROC (%): 68.64 -> 70.85 | No clear effect |
| feature-based vs confidence-based methods: mean AUROC drop between similar and dissimilar artefacts | feature-based methods · ISIC real benchmarks | mean AUROC drop (pp): 7.1 vs 1.5 | Hurt |
| train on hypointense-heart counterfactuals instead of hyperintense (ROI intensity flip) (ablation) | Mahalanobis · CheXpert synthetic squares | AUROC drop (%): 6.31 -> 4.28 (preference reverses) | Mixed |
| swap artefact colour with synthetic counterfactual (dissimilar) for colour charts | Mahalanobis · ISIC colour charts | AUROC (%): 96.11 -> 93.77 | Hurt |
| add light colour jitter augmentation (ablation) | KNN · ISIC ink | AUROC (%) on dissimilar ink: 70.10 -> 77.3 | Improved |
| add colour jitter augmentation (ablation) | DICE · ISIC ink | AUROC (%) on similar ink: 69.04 -> 61.8 | Hurt |
| use heavy colour jitter instead of none (ablation) | ResNet18 primary model · ISIC | ID accuracy (pp): -5.5 | Hurt |
| add nuisance subspace projection (ablation) | Mahalanobis · ISIC ink | AUROC (%) on dissimilar ink: 63.6 -> 75.8 | Improved |
| add nuisance subspace projection (ablation) | FeatureNorm · ISIC ink | AUROC (%) on dissimilar ink: 52.9 -> 74.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Mahalanobis — Mahalanobis++: Improving OOD Detection via Feature Normalization (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.88)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.88)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.87)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.87)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · 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.
