Physical World & Real-World ImpactChaoyang University of Technology (Chayanon Sub-r-pa, Rung-Ching Chen)
YOLO-World v2 applied to MVTec AD with an extreme 1:9 train-validation split
Applied YOLO-World v2 to MVTec AD with an extreme 1:9 train-validation split — an average of 11.8 defect instances per category. By the authors' measurement, it beat YOLOv11s and YOLOv26s in 12 of 15 categories; the weakest category, with 9 training instances, reached 64.9% mAP@50
Why it matters
An experimental design that fits the OSAT reality, where a new defect class has to start from 10 images or fewer. However, the absolute performance is at the level of inspection support and triage, not production pass/fail decisions
Source — Primary sources first; secondary reports are labelled “Press”
Officialmdpi.comhttps://www.mdpi.com/1999-4893/19/4/259Added in Week 35, 2026 · See that week