Research map / Label-efficient and robust learning / OOD and anomaly detection
Improving Out-of-Distribution Detection via Dynamic Covariance Calibration
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 dynamic covariance calibration for test-time OOD detection, adjusting the within-class covariance of a Mahalanobis-style distance with real-time input features restricted to the residual space.
Problem
Static prior geometry (Mahalanobis, subspace methods) estimated from training data is distorted by ill-distributed training samples, reducing OOD detection sensitivity in test-time settings.
Approach
Dynamically shrink the prior covariance along the real-time input feature direction (Sigma - f f^T), restricted to the residual subspace of the within-class covariance, so the distance metric adapts to each test input without retraining.
Key results
- On CIFAR-100 with DenseNet, the method reaches 92.38 AUROC and 29.98 FPR95 averaged over six OOD datasets, versus 89.25 / 50.57 for FDBD.
- On ImageNet-1k average over four models, it reports 88.69 AUROC and 49.72 FPR95, improving ResNet-50 to 87.43 AUROC versus 84.47 for FDBD.
- On DINO, it reaches 91.65 AUROC and 38.23 FPR95, slightly above Mahalanobis (91.57 / 39.27).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dynamic covariance calibration instead of the strongest reported average baseline | CIFAR DenseNet OOD detection · CIFAR average over 6 OOD datasets | AUROC: 93.24 -> 94.61 | Helped |
| Use dynamic covariance calibration instead of the strongest reported average baseline | CIFAR WideResNet OOD detection · CIFAR average over 6 OOD datasets | FPR95: 47.22 -> 31.73 | Helped |
| Use dynamic covariance calibration instead of the strongest reported average baseline | ImageNet-1k OOD detection · ImageNet-1k average | FPR95: 53.53 -> 49.72 | Helped |
| Use dynamic covariance calibration instead of Mahalanobis distance | DINO OOD detection · DINO, average over 6 OOD datasets | FPR95: 39.27 -> 38.23 | Helped |
| Add dynamic matrix estimation to the normalized-feature Mahalanobis baseline (ablation) | OOD detection across CIFAR-10 DenseNet, CIFAR-100 DenseNet, ImageNet-1k ViT, and ImageNet-1k ResNet-50 · Four ablation settings | FPR95: 18.86 -> 17.51; 35.84 -> 36.61; 33.16 -> 32.31; 60.43 -> 61.77 | Mixed |
| Add residual space projection to DME (ablation) | OOD detection across CIFAR-10 DenseNet, CIFAR-100 DenseNet, ImageNet-1k ViT, and ImageNet-1k ResNet-50 · Four ablation settings | FPR95: 17.51 -> 23.92; 36.61 -> 53.70; 32.31 -> 86.32; 61.77 -> 52.17 | Mixed |
| Add dynamic covariance modeling to DME (ablation) | OOD detection across CIFAR-10 DenseNet, CIFAR-100 DenseNet, ImageNet-1k ViT, and ImageNet-1k ResNet-50 · Four ablation settings | FPR95: 17.51 -> 16.38; 36.61 -> 30.16; 32.31 -> 27.05; 61.77 -> 60.49 | Mixed |
| Add dynamic adjustment to Euclidean distance (ablation) | Euclidean-distance OOD detection · CIFAR-10 and CIFAR-100 DenseNet; ImageNet-1k ViT | AUROC: 96.429 -> 96.443; 89.217 -> 89.563; 93.267 -> 93.993 | Helped |
| Add dynamic adjustment to Euclidean distance (ablation) | Euclidean-distance OOD detection · ImageNet-1k ResNet-50 | AUROC: 87.794 -> 87.794 | No clear effect |
| Add dynamic adjustment to Euclidean distance (ablation) | Euclidean-distance OOD detection · ImageNet-1k DINO | AUROC: 88.095 -> 89 | Helped |
| Add dynamic adjustment to Euclidean distance (ablation) | Euclidean-distance OOD detection · CIFAR-10 and CIFAR-100 DenseNet; ImageNet-1k ViT, ResNet-50, and DINO | FPR95: 18.101 -> 18.023; 41.032 -> 39.836; 32.817 -> 28.691; 46.163 -> 46.162; 49.891 -> 46.425 | Helped |
Surprising findings
- Euclidean-distance variant gains are marginal, suggesting that principal-space geometry dominates performance once adjustment is applied.
- Subspace methods (ViM, Neco, WDiscOOD) perform weakly on ResNet-50, which the authors attribute to residual-space dependence on feature quality.
Limitations
- Computational complexity is O(n^3) in feature dimension due to matrix inversion per input.
- Gains on Euclidean-distance variant are insignificant, and performance on some settings (e.g., ImageNet ViT and DeiT FPR95, CIFAR-100 ImageNet Swin) is not best.
- Not evaluated on vision-language or large language models; the authors acknowledge this limitation.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-1k SVHN Textures Places365 LSUN LSUN_Resize iSUN iNaturalist ImageNet-O OpenImage-O
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-1k OpenOOD
Metrics: AUROC FPR95
Baselines: MSP Energy maxLogit ODIN Mahalanobis GEM KNN ReAct DICE Line FDBD VIM
Builds on: Mahalanobis distance OOD detection ViM (virtual logit matching) NECO FDBD WDiscOOD residual space method
Closest papers by meaning
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · similarity 0.88)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.87)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.87)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.87)
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · 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.