Tech notes · 3
Tech notes
What the lab learned on the line — algorithms, performance work and installation know-how.
2026-08-041 views2026-07-222 views2026-07-221 views
The bottleneck was the PNG encoder — a 16× faster augmentation
Over 90% of augmentation time went into PIL's PNG encoder. One compression level and parallel saving took each image from 4.3 s to 0.26 s.
Aligning parts that are not round — moments and principal axes
Training is more stable when every part faces the same way, round or not. How first-, second- and third-order moments give position, orientation and front/back.
Straighten first, then judge — Multi-Inspector's yolo_align mode
A new mode that runs YOLO on images that are only rotated and centered, without Canny preprocessing — reproducing training-time alignment on the line.