# before
import motmetrics as mm

# after: same code, motrics underneath
import motrics.compat.motmetrics as mm

acc = mm.MOTAccumulator(auto_id=True)
for gt_ids, gt_boxes, pred_ids, pred_boxes in sequence:
    dists = mm.distances.iou_matrix(gt_boxes, pred_boxes, max_iou=0.5)
    acc.update(gt_ids, pred_ids, dists)

summary = mm.metrics.create().compute(acc, metrics=mm.metrics.SUPPORTED, name="acc")

pip install motrics[compat] (pulls in pandas, needed only for this subpackage).

The full motmetrics.metrics.motchallenge_metrics field set is supported: mota, motp, idf1, idp, idr, recall, precision, num_false_positives, num_misses, num_switches, num_unique_objects, mostly_tracked, partially_tracked, mostly_lost, num_fragmentations, num_transfer, num_ascend, num_migrate.

See python/motrics/compat/motmetrics/ for what else differs (e.g. no events/mot_events DataFrame).

# before
import trackeval

# after: same code, motrics underneath
import motrics.compat.trackeval as trackeval

eval_config = trackeval.Evaluator.get_default_eval_config()
evaluator = trackeval.Evaluator(eval_config)

dataset_config = trackeval.datasets.MotChallenge2DBox.get_default_dataset_config()
dataset_config["GT_FOLDER"] = "data/gt/mot_challenge/"
dataset_config["TRACKERS_FOLDER"] = "data/trackers/mot_challenge/"
dataset_list = [trackeval.datasets.MotChallenge2DBox(dataset_config)]

metrics_list = [trackeval.metrics.HOTA(), trackeval.metrics.CLEAR(), trackeval.metrics.Identity()]

results, messages = evaluator.evaluate(dataset_list, metrics_list)
print(results["MotChallenge2DBox"]["my_tracker"]["COMBINED_SEQ"]["pedestrian"]["CLEAR"]["MOTA"])

Same class names, config keys, directory/seqmap conventions, and result shape as real TrackEval. No trackeval/scipy install required, only numpy (a core dependency already).

See python/motrics/compat/trackeval/ for the full list of what differs from real TrackEval.

Using motrics' own API directly (faster than the compat layer, no per-frame Python bookkeeping)? Here's how the field names line up:

• Project scaffolding (build, lint, packaging, CI)
• Published to PyPI, automated tag-and-release on every Cargo.toml version bump
• Box IoU + assignment (Hungarian/greedy) primitives
• CLEAR metrics: MOTA/MOTP, ID switches, FP/FN, MT/PT/ML, fragmentations, and the derived MODA/sMOTA/MOTAL/CLR_Re/CLR_Pr fields
• Identity metrics (IDF1/IDP/IDR)
• HOTA (DetA, AssA, alpha sweep)
• MOTChallenge ingest, integration tests, TrackEval numeric parity tests
• Benchmark & parity infrastructure vs TrackEval and py-motmetrics on real MOTChallenge data, validated in CI
• Zero-copy NumPy input path; xyxy/xywh box formats
• Precomputed-similarity core inputs (compute_*_from_similarity)
• motrics.compat.motmetrics: drop-in MOTAccumulator, full motchallenge_metrics field set including switch subtypes (num_transfer/num_ascend/num_migrate)
• motrics.compat.trackeval: drop-in Evaluator/MotChallenge2DBox/ HOTA/CLEAR/Identity
• Ergonomic native API: Frames, evaluate(), streaming Accumulator for CLEAR + Identity
• Mask-IoU similarity kernel (RLE codec, ignore-region/IoA semantics) for mask-based adapters
• 3D IoU similarity kernel (oriented boxes) for KITTI-3D
• Dataset adapters: DanceTrack, KITTI 2D-box, KITTI-MOTS, DAVIS, BDD100K, KITTI-3D (see Datasets)
• Other TrackEval metrics: IDEucl, TrackMAP, VACE, JAndF (DAVIS's native metric)
• Remaining compat.trackeval Evaluator behaviors: parallel evaluation, printing/plotting, zipped input, DO_PREPROC=False, MOT15, BREAK_ON_ERROR