Reproduce “Real-IAD Variety: Pushing Industrial Anomaly Detection Dataset to a Modern Era” on OpenResearch

About this paper

Wenbing Zhu, Chengjie Wang, Bin-Bin Gao, Jiangning Zhang, Guannan Jiang, Jie Hu, Zhenye Gan, Lidong Wang and 7 more

Real-IAD Variety introduces a five-view industrial anomaly-detection benchmark with 198,950 images across 160 categories and pixel-level defect annotations. Across category-count subsets, UniAD image-level AUROC falls from 78.5% to 67.5%, while zero-shot AdaptCLIP changes from 73.2% to 72.9%. With four normal prompts, AdaptCLIP reaches 88.1%. These results show method-dependent scaling, not isolated category-count effects or proven deployed performance.

Key insights

• Traditional multi-class unsupervised anomaly detection (MUAD) methods exhibit substantial performance degradation as the number of object categories increases on the Real-IAD Variety dataset.
• Zero-shot and few-shot anomaly detection (ZSAD/FSAD) methods, particularly those based on Vision-Language Models (VLMs), demonstrate superior stability and generalization across a high diversity of categories.
• Pre-training foundational anomaly detection models on the Real-IAD Variety dataset significantly enhances their few-shot generalization performance on external IAD benchmarks.

Summary by alphaXiv.