What it does

This is a reference implementation for 14 object-detection metrics—mAP, AP@50, AP@[.5:.05:.95], AR variants, and the more exotic Spatio-Temporal Tube AP—wrapped in a visual interface. It reads more than 8 annotation formats (COCO, PASCAL VOC, YOLO, Labelme, CVAT, etc.) without forcing you to convert your detections to XML or JSON first. The authors published an accompanying paper in Electronics (2021) to document how small implementation differences in interpolation or IOU thresholds can skew reported results.

The interesting bit

The README spends real effort on the boring part: a step-by-step toy example showing how 11-point interpolation versus all-point interpolation changes your AP by a percentage point or two, and how stricter IOU thresholds turn the same detections into different precision-recall curves. That pedantry is the value—most papers quietly use incompatible metric variants and call it mAP.

Key highlights

• Supports 14 metrics including COCO-style AP scales and video-specific STT-AP
• Reads 8+ bounding-box formats natively (no conversion pipeline needed)
• Visual GUI plus command-line usage for both images and spatio-temporal tubes
• Ships with a 12-image worked example and precision-recall plots
• Peer-reviewed paper and BibTeX citation provided

Caveats

• Version badge shows 0.1, suggesting early or slow-moving releases
• README is thorough but long; the “How to use” section is brief and leaves setup details to the reader
• Travis CI badge is present; current build status is unclear from the sources

Verdict

Worth bookmarking if you evaluate object detectors across multiple datasets or need to sanity-check why your mAP differs from a paper’s. Skip it if you already live entirely inside COCO’s official API and never touch PASCAL or YOLO formats.