Continual-MEGA: A Large-scale Benchmark for Generalizable Continual Anomaly Detection

TL;DR

Continual-MEGA introduces a large-scale benchmark for continual anomaly detection, integrating seven public datasets with the newly created ContinualAD to enable realistic continual learning and continual zero-shot evaluation. It proposes ADCT, a CLIP-based baseline that uses lightweight mixture-of-experts adapters and synthetic feature generation to balance continual adaptation with zero-shot generalization. Extensive experiments show substantial room for improvement in pixel-level localization, while ADCT and the ContinualAD dataset provide robust performance gains and enhanced generalization across CZSL settings. The work offers a practical benchmark and baseline that can drive development of more generalizable continual anomaly detection methods for real-world deployment.

Abstract

In this paper, we introduce a new benchmark for continual learning in anomaly detection, aimed at better reflecting real-world deployment scenarios. Our benchmark, Continual-MEGA, includes a large and diverse dataset that significantly expands existing evaluation settings by combining carefully curated existing datasets with our newly proposed dataset, ContinualAD. In addition to standard continual learning with expanded quantity, we propose a novel scenario that measures zero-shot generalization to unseen classes, those not observed during continual adaptation. This setting poses a new problem setting that continual adaptation also enhances zero-shot performance. We also present a unified baseline algorithm that improves robustness in few-shot detection and maintains strong generalization. Through extensive evaluations, we report three key findings: (1) existing methods show substantial room for improvement, particularly in pixel-level defect localization; (2) our proposed method consistently outperforms prior approaches; and (3) the newly introduced ContinualAD dataset enhances the performance of strong anomaly detection models. We release the benchmark and code in https://github.com/Continual-Mega/Continual-Mega.

Figures (11)

• Figure 1: Motivation for using Continual Learning (CL) and Continual Zero-Shot Learning (CZSL) in anomaly detection, enabling models to handle evolving defects over time and generalize to unseen anomalies without retraining.
• Figure 2: Illustration of statistics of various datasets. Each graph (from left to right) shows the number of classes, number of images, and pixel value variance for public datasets, as well as our proposed ContinualAD and Continual-MEGA. Image variance is defined as the average per-pixel variance for each class.
• Figure 3: Example visualizations of sample images from various public anomaly detection datasets and the proposed ContinualAD dataset. Green boxes indicate normal images, while red boxes represent anomaly images.
• Figure 4: Overview architecture of training process. Synthetic feature generation is performed exclusively during the training phase. At inference, adapters trained for each task are accumulated for evaluation. A detailed overview of the inference stage is provided in Appendix Figure \ref{['fig:overview_inference']}.
• Figure 5: Learning curves on Scenario 2 (10 classes per task). We plot the average of image-level AUROC and pixel-level AP over all tasks seen so far as the model is incrementally trained from Task 1 to Task 6. This plot highlights that under matched compute, ADCT (Ours) exhibits superior stability, preserving performance across tasks substantially better than MVFA, which corroborates our quantitative FM results in Table \ref{['tab:scenario2_full']}.