A textile anomaly model's lab metric may not remain unchanged under production conditions. Measure lighting flicker, fabric tension and lot changes, camera alignment, lens contamination and differences in defect definitions between annotators. Validate the data and imaging chain before attributing a score gap to the model.
Dataset discipline
Public datasets may not represent the random weave, camera setup and defect taxonomy of your line. Collect representative data across shifts, lots, colours, machine speeds and lighting conditions. Compare pixel masks with bounding boxes on the same acceptance set using precision, recall and annotation effort; annotation type alone does not guarantee an outcome.
Operator feedback loop
Add false-positive and false-negative examples to the feedback set only after operator confirmation and data-quality review. Fine-tuning or retraining should not follow a fixed weekly or monthly recipe; trigger it from drift metrics, sufficient new samples, error cost and release approval. Do not promote a new model without validating it on a separate acceptance set.
The lab-to-line gap is a diagnostic signal. Break metrics down by data slice, lot, shift and environmental condition; the corrective action may be lighting or mounting, but it may also be labels, the model or the decision threshold.