Validation Prediction Review

Scoring method: Each image is scored by its F1 score at IoU ≥ 0.5: F1 = 2·TP / (2·TP + FP + FN). A predicted box is a True Positive (TP) if it overlaps a ground-truth box of the same class with IoU ≥ 0.5 (greedy highest-IoU matching, at most one GT per prediction). Unmatched predictions are False Positives (FP); unmatched GT boxes are False Negatives (FN). Images with no GT boxes and no predictions receive F1 = 1.0 (vacuously perfect); images with predictions but no GT receive F1 = 0.0.

Best examples have F1 closest to 1 — the model correctly locates and classifies nearly all parts. Worst examples have F1 closest to 0 — many parts are missed or mislabelled.