Data for: Product-level false-alarm calibration for multi-view industrial inspection under illumination shift

Published: 19 November 2026| Version 2 | DOI: 10.17632/cp2rh99fm3.2
Contributor:
mengmeng zhang

Description

Authors: Mengmeng Zhang, Chenggang Xu, Zhouqi Kang, Xiaoxiao Cheng (corresponding). This dataset supports an analysis of product-level false-alarm calibration in multi-view industrial inspection under illumination shift. It contains 273,240 frozen image-level anomaly scores for PatchCore, reverse distillation and DINOv2-based PatchCore, a product-disjoint manifest for 998 products, and 500 stored normal-product re-splits per category. Additional files contain PatchCore target-illumination memory-bank scores, I02-reference scores, and alternative k-centre-start scores for Bird, Car and Tube. The updated package includes portable analysis code, current numerical summaries, category-bootstrap uncertainty estimates, exact sign-flip tests with BH and BY corrections, verification reports and the seven current manuscript images. The original frozen scores and split definitions are unchanged. The executable analyses reproduce results from the stored detector scores; they do not train models or run image inference. Original M2AD images are not redistributed and should be obtained from the dataset authors. Version 2 adds reference-light and bank-start robustness data, the revised 15 November statistical scripts and results, package-relative execution paths, documentation corrections and checksums. It is aligned with the current manuscript title.

Files

Categories

Computer Science

Licence