Research map / Label-efficient and robust learning / OOD and anomaly detection

InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class 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.

Introduces InCLAD, a task-incremental continual benchmark for semi-supervised (normal-only) industrial visual anomaly detection built from five public datasets, with curriculum and multi-dataset streams, evaluated at image and pixel level with ROC-AUC and BWT.

Problem

Industrial anomaly detection benchmarks are static, so they do not capture extending inspection systems to new components over time without catastrophic forgetting of earlier components; continual anomaly detection under normal-only training is underexplored.

Approach

Frame continual industrial anomaly detection as task-incremental, normal-only learning where each industrial component is a task, define difficulty-based curricula (easy-to-hard, hard-to-easy, random) from averaged single-task ROC-AUC ranks of PatchCore, PaDiM and CFA, and add a cross-dataset 27-task stream; adapt PatchCore, PaDiM and CFA to continual settings and evaluate nine methods.

Key results

• Continual learning raises average image-level ROC-AUC from 0.588 (NCL) to 0.737 (CL) and pixel-level ROC-AUC from 0.718 to 0.860 across single-dataset streams.
• BWT improves from -0.088 to -0.027 at image level and from -0.074 to -0.010 at pixel level under continual learning.
• On the 27-task MultiDataset stream, CL image-level ROC-AUC reaches about 0.79 versus near-chance NCL (~0.53-0.55).

Surprising findings

• Naive non-continual detectors on the cross-dataset stream are near chance at image level (ROC-AUC about 0.53-0.55).
• Task ordering affects image-level CL performance more than pixel-level performance (pixel CL ROC-AUC 0.849-0.870 across orderings).

Limitations

• Limited to existing public industrial datasets that only partially capture temporal, sensor, and operational variability.
• Continual variants rely on replay or method-specific adaptations; no standardized compute or memory budget is enforced.
• Single-dataset averages pool all nine methods, so method-specific rankings require per-method breakdowns.

Experimental setup

Datasets: MVTec AD VisA BTAD MPDD DAGM

Benchmarks: MVTec AD VisA BTAD MPDD DAGM InCLAD MultiDataset

Metrics: ROC-AUC backward transfer (BWT)

Baselines: GANomaly PatchCore PaDiM CFA STFPM RD4AD EfficientAD PaSTe FastFlow

Builds on: PatchCore PaDiM CFA

Closest papers by meaning

• Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt (AAAI 2024 · similarity 0.89)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.88)
• Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.87)
• Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.87)
• Co-Progression Knowledge Distillation with Knowledge Prototype for Industrial Anomaly Detection (AAAI 2025 · similarity 0.86)
• Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.86)