imaadd05/visa-anomaly-detection

VisA — Visual Anomaly Dataset

Mirror of the VisA (Visual Anomaly) dataset for research use. Staged as a proxy/pretraining dataset for CoRe's Situational Control paint-inspection work (core-lab/situational-control).

Source

• Original repo: https://github.com/amazon-science/spot-diff
• Original download: https://amazon-visual-anomaly.s3.us-west-2.amazonaws.com/VisA_20220922.tar
• License: CC BY 4.0 (confirmed in the source repo README)

Contents

10,821 images across 12 object subsets (candle, capsules, cashew, chewinggum, fryum, macaroni1, macaroni2, pcb1-4, pipe_fryum), 9,621 normal + 1,200 anomalous. Pixel-level segmentation masks included for anomalous images. Structure preserved as-is from the original tar:

<object>/Data/Images/{Normal,Anomaly}/*.JPG
<object>/Data/Masks/Anomaly/*.png
<object>/image_anno.csv

Citation

@article{zou2022spot,
  title={SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation},
  author={Zou, Yang and Jeong, Jongheon and Pemula, Latha and Zhang, Dongqing and Dabeer, Onkar},
  journal={arXiv preprint arXiv:2207.14315},
  year={2022}
}

Paper: "SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation", Zou et al., ECCV 2022.

All credit for data collection and annotation belongs to the original authors (Amazon Science). This is an unmodified mirror for convenience of access.