Comparing detectors without format gymnastics

What it does

This Python repo calculates core object detection metrics—Precision-Recall curves and Average Precision—using the same math as the PASCAL VOC challenge. It takes plain bounding boxes as input, sparing you the usual ritual of converting model outputs into XML or JSON. The authors state they compared their results against official evaluation scripts and found them identical.

The interesting bit

The project treats metric implementation as a standardization problem, not a coding exercise. By locking down details like 11-point interpolation and IOU thresholds, it prevents the small arithmetic divergences that quietly corrupt benchmarks across different papers. The included survey paper adds an unusual layer of academic authority to what is essentially a referee’s toolkit.

Key highlights

• Implements PASCAL VOC metrics (11-point and all-point interpolation) without dataset-specific format requirements.
• Accepts raw ground-truth and predicted bounding boxes directly.
• Claims exact parity with official competition implementations.
• Bundles a detailed explainer on IOU, precision, recall, and AP calculation.
• Supported by published peer-reviewed papers (2021 Electronics and a 2020 conference survey).

Caveats

• The README explicitly flags a newer successor repo that adds COCO metrics, a UI, and video evaluation; this version is missing those features.
• The documentation focuses heavily on PASCAL VOC; support for COCO’s full 12-metric suite is not demonstrated here.

Verdict

Use this if you want a battle-tested, no-frills reference implementation for PASCAL VOC-style mAP without rewriting evaluation scripts. If you need COCO metrics or a point-and-click interface, the authors themselves point to their newer tool.