hotcoco

A drop-in replacement for pycocotools, written in Rust. Same numbers, bit for bit. Up to 113× faster. No compiler to install.

pip install hotcoco

from hotcoco import init_as_pycocotools
init_as_pycocotools()

# Every pycocotools import now resolves to hotcoco. Nothing else changes.
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval

Or import it directly; the API is the same:

from hotcoco import COCO, COCOeval

coco_gt = COCO("instances_val2017.json")
coco_dt = coco_gt.load_res("detections.json")

ev = COCOeval(coco_gt, coco_dt, "bbox")
ev.run()

Documentation | Changelog | Roadmap

Why switch

• The numbers do not change. The precision, recall, and scores arrays and every summary metric are bit-identical to pycocotools' on COCO val2017, for bbox, segm, and keypoints. Parity is a hard requirement for every change; CONTRIBUTING.md describes how it is verified.
• It is fast. The benchmarks compare hotcoco, pycocotools, and three other evaluators on time and peak memory for bbox, segm, and keypoints on COCO val2017, with the hardware and the method.
• It installs anywhere. Prebuilt wheels for Linux, macOS, and Windows. No Cython, no C compiler, no Microsoft Build Tools.

Beyond pycocotools

The same engine covers more than COCO detection. Each item links to its guide.

• LVIS and Open Images protocols, and oriented boxes.
• Panoptic quality, matching panopticapi on val2017, and semantic segmentation, matching mmsegmentation.
• Streaming evaluation inside a validation loop, straight from the detector's prediction array, with shards that merge across ranks.
• TIDE error analysis, confusion matrices, calibration, label-error detection, and model comparison with bootstrap confidence intervals.
• Metric functions on plain arrays, the way sklearn.metrics works, for building your own metric.
• A dataset browser, a healthcheck, and format conversion to and from YOLO, Pascal VOC, CVAT, DOTA, and Open Images CSV.

The full list is on the documentation site.

Contributing

Contributions are welcome. See CONTRIBUTING.md for the architecture overview, build and test workflow, and pre-commit checks.

License

MIT