Answers are grounded in this paper and cite their sources.
Garment appearance inspection is often formulated as axis-aligned box detection and ranked by COCO mAP@0.5:0.95. Sewn-garment defects vary in scale, aspect ratio, and edge contrast within the same class. Intraclass variation does not preclude high intersection-over-union (IoU) for an individual sample: IoU measures agreement with that sample's annotated box. This study evaluates a 19-class corpus of 14,352 VGA garment-defect images using production-report-grouped splits and reports box-level detection alongside image-level multilabel recognition. Baseline YOLO26n and the proposed training configuration are each evaluated with three random seeds. The proposed configuration changes task-aligned assignment, box regression and image-level auxiliary supervision. Ablation results identify reducing the task-aligned-assignment exponent β as the largest isolated source of the observed gains, while the complete configuration provides only a modest additional improvement over the reduced-β-only setting. The evaluated configuration improves several detection metrics under the scale and aspect-ratio variation represented in this dataset, although gains vary across classes and are concentrated mainly in the reduced task-alignment exponent. Performance is reported at multiple IoU thresholds alongside image-level metrics to characterize complementary detection and recognition criteria.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: