Research map / Label-efficient and robust learning / OOD and anomaly detection
A Geometry-Based View of Mahalanobis OOD Detection
Denis Janiak, Jakub Binkowski, Tomasz Jan Kajdanowicz
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 |
|---|---|---|---|
| select the radial exponent beta from ID geometry (RS-MD) instead of fixed beta=0 (ablation) | MD · ImageNet-1K ID vs 5 OpenOOD datasets, average over 12 backbones | FPR@95: 37.2 -> 35.3 | Improved |
| select beta from ID geometry (RS-MD) instead of fixed beta=1 (L2 normalization) (ablation) | MD++ · ImageNet-1K ID vs 5 OpenOOD datasets, average over 12 backbones | FPR@95: 36.0 -> 35.3 | Improved |
| L2-normalize features (beta=1) (ablation) | MD · ImageNet-1K ID vs 5 OpenOOD datasets, average over 12 backbones | FPR@95: 37.2 -> 36.0 | Improved |
| select beta from ID geometry (RS-RMD) instead of beta=0 (ablation) | RMD · ImageNet-1K ID vs 5 OpenOOD datasets, average over 12 backbones | FPR@95: 37.2 -> 35.8 | Improved |
| select beta from ID geometry (RS-RMD) instead of beta=1 (ablation) | RMD++ · ImageNet-1K ID vs 5 OpenOOD datasets, average over 12 backbones | FPR@95: 36.1 -> 35.8 | Improved |
| proxy-selected beta (RS-MD) instead of beta=0 (ablation) | MD · ViT In1k | FPR@95: 45.7 -> 45.5 | No clear effect |
| proxy-selected beta (RS-MD) instead of beta=0 (ablation) | MD · ViT In21k->In1k | FPR@95: 35.7 -> 35.8 | No clear effect |
| use proxy-selected beta for RMD instead of beta=0 (ablation) | RMD · EVA02 In1k | FPR@95: 40.3 -> 39.6 | Improved |
| use RS-MD instead of MD++ (ablation) | MD++ · DeiT3 In1k | FPR@95: 43.0 -> 43.2 | No clear effect |
| use RMD instead of MD on pretrained-only (not ImageNet fine-tuned) models | MD · EVA02-In21k, ViT-In21k | FPR@95 | Improved |
| swap MD for RMD on DeiT3 In1k | MD · DeiT3 In1k | FPR@95: 43.3 -> 40.8 | Improved |
| swap MD for RMD on EVA02 In1k | MD · EVA02 In1k | FPR@95: 37.6 -> 40.3 | Hurt |
Closest papers by meaning
- Mahalanobis++: Improving OOD Detection via Feature Normalization (ICML 2025 · similarity 0.92)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.89)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.89)
- Improving Out-of-Distribution Detection via Dynamic Covariance Calibration (ICML 2025 · similarity 0.87)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.87)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 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.
