Answers are grounded in this paper and cite their sources.
Evaluating manufacturing vision requires strict isolation of repeated acquisitions between training and test data. We analyzed 1,120 nonpublic welding inspection images from two quality classes and 140 acquisition groups. After removing exact duplicates, we varied unique frames per group from one to five and compared image-random, acquisition-group, and paired-group five-fold validations across 30 assignments. Same-group exposure under image-random splitting reached 99.9%. Consequently, the balanced accuracy of HOG-linear SVM increased from 0.376 to 0.762 under image-random splitting but decreased from 0.409 to 0.288 under acquisition-group splitting, while paired-group performance remained at ~0.55. A compact convolutional baseline showed a much smaller but directionally consistent trend, with balanced accuracies of 0.524 (image random) and 0.484 (acquisition group). Decision-margin directions changed with the split unit, thus rendering model superiority unstable. Therefore, repeated frames amplify evaluation bias without adding independent cases. Performance reports should quantify exposure and align split units with intended deployment cases.
No takes yet. Share an insight, caveat, or question.
Yun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: