Drop ground truth and predictions
YOLO: gt/*.txt and pred/*.txt where predictions have a sixth confidence column. COCO: annotations.json and results.json.
Fine-tuning
mAP Calculator for Object Detection
Compare detectors on your own data without a framework. Drop ground-truth labels and predictions (YOLO txt with confidence, or COCO results JSON), and the tool matches boxes greedily by IoU per class, builds precision-recall curves, and reports AP per class at IoU 0.5 and averaged over 0.5 to 0.95, plus precision and recall at a chosen confidence threshold.
Loading the tool…
How it works
Choose thresholds
IoU 0.5 for the classic metric; the 0.5:0.95 average follows the COCO convention. Confidence threshold for the operating-point precision and recall.
Read and export
AP table per class, PR curves, and a JSON report you can commit next to the model.
Formats, limits and the numbers that matter
Matching
Predictions are sorted by confidence; each matches the highest-IoU unmatched ground truth of the same class above the threshold, exactly as in the VOC and COCO evaluators.
AP integration
101-point interpolation (COCO) by default; the 11-point VOC2007 method is an option for comparison with older papers.
Small, medium, large
With image sizes available, AP is also reported per COCO size bucket.
Common questions
Why is my mAP lower than the trainer's?
Trainers evaluate at a low confidence threshold (0.001) to fill the PR curve; if your prediction file was exported at 0.25, the curve is truncated. Export predictions at the lowest confidence you can.
Segmentation mAP?
Boxes only here; mask IoU is planned.
How many images are needed?
Per class, a few hundred ground-truth boxes for a stable number; the report prints the count per class.