Research map / Label-efficient and robust learning / OOD and anomaly detection
CADRef: Robust Out-of-Distribution Detection via Class-Aware Decoupled Relative Feature Leveraging
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 CARef (class-aware relative feature error) and CADRef (sign-decoupled positive/negative relative error scaled by logit scores) for post-hoc OOD detection on ImageNet-1k and CIFAR benchmarks.
Problem
Post-hoc OOD detection for DNNs, where logit-based and feature-based methods neglect information inherent in sample features, and feature-shaping methods show architecture-specific collapse.
Approach
Measure the normalized l1 error between a sample's penultimate feature and its class-average training feature, then decouple this error into positive and negative parts by the sign alignment with the max-logit weights and rescale the positive part by a logit score (Energy by default).
Key results
- CADRef improves average AUROC over the best baseline by 3.27% and reduces FPR95 by 6.32% on ImageNet-1k across seven architectures.
- CARef improves average AUROC by 2.82% and reduces FPR95 by 4.03% on ImageNet-1k.
- CADRef reaches 88.19 average AUROC on ImageNet-1k versus 84.92 for OptFS, the best baseline in average AUROC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the strongest baseline with CARef | OOD detection on ImageNet-1k · ImageNet-1k (average across models) | AUROC: 2.82% | Helped |
| Replace the strongest baseline with CADRef | OOD detection on ImageNet-1k · ImageNet-1k (average across models) | AUROC: 3.27% | Helped |
| Replace the strongest baseline with CARef | OOD detection on ImageNet-1k · ImageNet-1k (average across models) | FPR95: 4.03% | Helped |
| Replace the strongest baseline with CADRef | OOD detection on ImageNet-1k · ImageNet-1k (average across models) | FPR95: 6.32% | Helped |
| Use normalized relative feature error instead of ℓ1 distance or feature ℓ1 norm alone (ablation) | CARef · ImageNet-1k (average across models) | AUROC: 62.19; 40.85 -> 87.74 | Helped |
| Add negative error to positive error without error scaling (ablation) | CADRef · ImageNet-1k (average across models) | AUROC: 82.52 -> 87.70 | Helped |
| Add error scaling to the positive error (ablation) | CADRef · ImageNet-1k (average across models) | AUROC: 82.52 -> 87.58 | Helped |
| Add both scaled positive error and negative error (ablation) | CADRef · ImageNet-1k (average across models) | AUROC: 87.74 -> 88.19 | Helped |
| Add both scaled positive error and negative error (ablation) | CADRef · ImageNet-1k (average across models) | FPR95: 52.59 -> 50.30 | Helped |
| Add the Energy-based logit component to CARef | CADRef · ImageNet-O | AUROC: 78.03 -> 75.29 | Hurt |
| Add the Energy-based logit component to CARef | CADRef · SSB-hard | AUROC: 72.32 -> 74.58 | Helped |
| Add the Energy-based logit component to CARef | CADRef · Ninco | AUROC: 83.65 -> 85.36 | Helped |
Surprising findings
- Feature-shaping method ASH-S collapses below 50% AUROC on ViT-B/16, Swin-B, and ConvNeXt-B despite strong results on CNNs.
- Using the l1-norm of features as an OOD score (FeatureNorm-like) reverses on transformer models, giving AUROC below 30%.
Limitations
- CADRef's logit-based component fails on extremely hard OOD sets such as ImageNet-O, where Energy AUROC is near 50%.
- CARef degrades on CIFAR benchmarks, attributed to reduced feature dimensionality.
- Using MSP as the logit scale for CADRef substantially lowers performance, which is not explained.
Experimental setup
Datasets: ImageNet-1k CIFAR-10 CIFAR-100
Benchmarks: ImageNet-1k OOD (iNaturalist, SUN, Places, Texture, OpenImage-O, ImageNet-O) CIFAR-10/100 OOD (SVHN, LSUN-C, LSUN-R, iSUN, Textures, Places) ImageNet-O SSB-hard Ninco
Metrics: AUROC FPR95
Baselines: MSP MaxLogit ODIN Energy GEN ReAct DICE ASH-S OptFS ViM
Builds on: Energy GEN ViM ReAct ASH
Closest papers by meaning
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.89)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.88)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.88)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
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.