Research map / Label-efficient and robust learning / OOD and anomaly detection
Long-Tailed Anomaly Detection with Learnable Class Names
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.
Formulates long-tailed unsupervised anomaly detection with resampled MVTec, VisA and DAGM splits and proposes LTAD, which combines reconstruction-based and semantic AD using learned pseudo class names and a VAE for feature augmentation.
Problem
Unsupervised anomaly detection and localization across many object classes without class-name priors, when training data follows long-tailed class distributions and no defect images are available.
Approach
Combines transformer-based feature reconstruction with a semantic normal/abnormal classifier in ALIGN text space, where class names are replaced by learned pseudo class-name prompts, and uses a text-conditioned VAE to synthesize tail-class features for augmentation.
Key results
- On MVTec exp100, LTAD reaches 88.86 image AUROC vs 87.70 for UniAD (best baseline in that config).
- LTAD outperforms baselines on 29 of 32 tasks across the three datasets.
- On DAGM, LTAD beats baselines by at least 9.1% (AD) and 5.9% (AS) across imbalance settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the semantic AD module to LTAD (ablation) | LTAD · MVTec step100 | detection AUROC: 87.05 -> 87.36 | Helped |
| Add the semantic AD module to LTAD (ablation) | LTAD · MVTec step100 | segmentation AUROC: 93.13 -> 93.83 | Helped |
| Replace the text-conditioned VAE with an autoencoder (ablation) | LTAD without SAD on MVTec step100 · MVTec step100 | detection AUROC: 87.05 -> 84.48 | Hurt |
| Replace the text-conditioned VAE with an autoencoder (ablation) | LTAD without SAD on MVTec step100 · MVTec step100 | segmentation AUROC: 93.13 -> 92.53 | Hurt |
| Use distribution-aware sampling instead of pc = 0.5 (ablation) | LTAD without SAD on MVTec step100 · MVTec step100 | detection AUROC: 87.05 -> 86.06 | Hurt |
| Initialize learned pseudo class names with dataset class names (ablation) | LTAD without SAD on MVTec step100 · MVTec step100 | detection AUROC: 87.05 -> 86.12 | Hurt |
| Add layer-aware projection (ablation) | LTAD semantic AD module without RM on MVTec step100 · MVTec step100 | segmentation AUROC: 87.18 -> 91.36 | Helped |
| Use max pooling instead of mean pooling across projected layers (ablation) | LTAD semantic AD module without RM on MVTec step100 · MVTec step100 | detection AUROC: 81.25 -> 84.12 | Helped |
| Use multiple normal/abnormal text prompts instead of a single prompt (ablation) | LTAD semantic AD module without RM on MVTec step100 · MVTec step100 | detection AUROC: 84.12 -> 77.19 | Hurt |
| Use the ALIGN text encoder and its class prior (ablation) | LTAD semantic AD module on MVTec step100 · MVTec step100 | detection AUROC: 59.79 -> 84.12 | Helped |
| Keep each learned pseudo class name assigned to its corresponding class (ablation) | LTAD semantic AD module on MVTec step100 · MVTec step100 | detection AUROC: 72.76 -> 84.12 | Helped |
| Use normal prompt “a” and abnormal prompt “a broken” (ablation) | LTAD semantic AD module · MVTec step100 | detection AUROC: 84.12 / 91.36 | Helped |
Surprising findings
- Using dataset class names as prompts underperforms learned pseudo class names, indicating class names are not always informative.
- Adding the semantic AD module gives gains in 31 of 32 tasks but the RegAD baseline beats LTAD on a few MVTec tasks.
Limitations
- Semantic AD relies on the pretrained ALIGN text encoder and the choice of normal/abnormal prompts strongly affects performance.
- Evaluated only on three industrial AD datasets with simulated long-tailed splits via resampling.
Experimental setup
Datasets: MVTec VisA DAGM
Benchmarks: MVTec VisA DAGM
Metrics: AUROC (image) AUROC (pixel) average/High/Low AUROC
Baselines: Cut & Paste MKD DRAEM RegAD UniAD AnomalyGPT
Builds on: UniAD ALIGN WinCLIP RegAD CoOp
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection (CVPR 2025 · similarity 0.88)
- Anomaly Heterogeneity Learning for Open-set Supervised Anomaly Detection (CVPR 2024 · similarity 0.88)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.87)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
- Weakly Supervised Anomaly Detection via Dual-Tailed Kernel (ICML 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.