VisA benchmark
VisA provides normal and anomalous product images with anomaly segmentation masks. PCB1, capsules and cashew are included. The official one-class split fits on normal images; labelled normal and anomalous images are used for evaluation. Multiclass masks can be reduced to binary anomaly masks. Normal images have no stored anomaly mask. Image-level labels establish benchmark anomaly status, not a validated diagnosis for a newly uploaded photograph.
Scope and attribution
The VisA dataset is released by Amazon Science under CC BY 4.0. Attribute Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang and Onkar Dabeer, SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation, ECCV 2022. DefectRAG sample previews are resized JPEG derivatives. Dataset annotations are benchmark evidence, not factory acceptance tolerances.