Research map / Label-efficient and robust learning / OOD and anomaly detection
ResAD: A Simple Framework for Class Generalizable Anomaly Detection
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 ResAD, a class-generalizable anomaly detector that learns residual features (feature minus nearest normal reference feature), constrains them to a hypersphere, and models their distribution with a normalizing flow so one model transfers to unseen classes without fine-tuning.
Problem
Anomaly detection models trained one-for-one or one-for-many degrade sharply on new object classes, and retraining or fine-tuning on target classes is often infeasible; the goal is a single model that generalizes to unseen classes using only few-shot normal references.
Approach
Learn the distribution of residual features (each patch minus its nearest few-shot normal reference patch), which varies less across classes than initial features; a shallow Conv+BN+ReLU Feature Constraintor with an abnormal-invariant OCC loss maps normal residuals onto a hypersphere, and a normalizing flow estimates their density for anomaly scoring.
Key results
- Under 4-shot cross-dataset evaluation, ResAD with WideResNet50 averages 89.7/97.3 image/pixel AUROC on industrial datasets, outperforming the best non-CLIP competitor RegAD (79.0/94.9).
- ResAD with ImageBind-Huge (ResAD†) reaches 88.0/96.3 average across all datasets at 4-shot, above WinCLIP (81.5/94.0) and InCTRL (81.5/-).
- Removing residual feature learning drops VisA-to-MVTecAD image AUROC from 90.5 to 72.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ImageBind-Huge instead of WideResNet50 as the feature extractor | ResAD · All Average, 4-shot | Image-level AUROC: 83.0 -> 88.0 | Helped |
| Add residual feature learning (ablation) | ResAD · MVTecAD, 4-shot, VisA to MVTecAD | Image-level AUROC: 72.8 -> 90.5 | Helped |
| Add Feature Constraintor (ablation) | ResAD · MVTecAD, 4-shot, VisA to MVTecAD | Image-level AUROC: 82.3 -> 90.5 | Helped |
| Add abnormal invariant OCC loss (ablation) | ResAD · MVTecAD, 4-shot, VisA to MVTecAD | Image-level AUROC: 84.9 -> 90.5; 5.6% | Helped |
| Use simple Conv+BN+ReLU rather than a more complex constraintor (ablation) | ResAD · MVTecAD, 4-shot, VisA to MVTecAD | Image-level AUROC: 87.6 -> 90.5; 86.0 -> 90.5; 84.1 -> 90.5; 82.3 -> 90.5; 81.0 -> 90.5 | Helped |
| Add residual feature learning to UniAD | UniAD · MVTecAD | Image-level AUROC: 67.4 -> 93.0; +25.6 | Helped |
| Add residual feature learning to UniAD | UniAD · VisA | Image-level AUROC: 52.1 -> 72.7; +20.4 | Helped |
| Add residual feature learning to UniAD | UniAD · BTAD | Image-level AUROC: 67.1 -> 87.3; +20.0 | Helped |
| Add residual feature learning to UniAD | UniAD · MVTec3D | Image-level AUROC: 51.7 -> 76.7; +25.0 | Helped |
| Increase few-shot normal references from 2 to 4 | ResAD · All Average | Image-level AUROC: 78.0 -> 83.0 | Helped |
| Use λ=0.1 rather than λ=1 for the abnormal-boundary loss (ablation) | ResAD · VisA to MVTecAD, 4-shot | Image-level AUROC: 90.5 -> 89.1 | Hurt |
| Use λ=0.5 rather than λ=1 for the abnormal-boundary loss (ablation) | ResAD · VisA to MVTecAD, 4-shot | Image-level AUROC: 90.5 -> 90.0 | Hurt |
Surprising findings
- The Feature Constraintor performs best as a single Conv+BN+ReLU layer; more complex constraintors (BottleNeck, multi-scale fusion) substantially degrade cross-class performance.
Limitations
- Experiments are limited to image modality; video and time series not thoroughly validated.
- Performance depends on representativeness of randomly selected few-shot reference samples; random selection may be insufficient for hard classes.
- Training with summed losses was unstable, requiring detaching features between the Feature Constraintor and the normalizing flow.
Experimental setup
Datasets: MVTecAD VisA BTAD MVTec3D BraTS ShanghaiTech
Benchmarks: MVTecAD VisA BTAD MVTec3D BraTS ShanghaiTech
Metrics: AUROC (image-level) AUROC (pixel-level)
Baselines: RDAD UniAD SPADE PaDiM PatchCore RegAD WinCLIP InCTRL
Builds on: normalizing flow BGAD semi-push-pull loss PatchCore one-class classification RealNVP
Closest papers by meaning
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.90)
- Anomaly Heterogeneity Learning for Open-set Supervised Anomaly Detection (CVPR 2024 · similarity 0.90)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.89)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.89)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.89)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 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.