fix(io): reject duplicate per-frame track ids

[AlexBodner]

Problem

CLEAR, HOTA and Identity all accumulate per-id counts with fancy-indexed +=, which NumPy buffers instead of adding:

counts = np.zeros(3)
counts[np.array([1, 1, 2])] += 1  # [0, 1, 1] — index 1 was written once, not twice

A track id repeated inside one frame therefore deflates the very totals the metrics divide by, and every reported number comes back quietly wrong.

Fix

_prepare_mot_sequence rejects a repeated id. TrackEval raises on the same input, so this matches the reference.

The check runs after filtering, so the two legitimate cases still pass:

• ids on rows dropped as ignored or non-pedestrian
• unconfirmed detections, which all share id -1 and are expected to repeat

Tests

• Duplicate ground-truth ids and duplicate tracker ids in one frame both raise.
• The same id across consecutive frames is normal tracking and still works.
• Duplicate ids on filtered rows and repeated -1 ids are allowed.

tests/io, tests/eval and tests/tune pass at 138.

Notes

Guarding in sequence preparation covers all three metric families at one chokepoint rather than patching each accumulation site in clear.py, hota.py and identity.py. Switching those to np.add.at would make duplicates count correctly instead of rejecting them, but duplicate ids in a frame are malformed input, and TrackEval treats them as an error too.

Made with Cursor

[Copilot]

Pull request overview

Rejects malformed MOT input containing duplicate per-frame track IDs before metric evaluation.

Changes:

• Adds per-frame uniqueness validation after relevant filtering.
• Adds coverage for duplicates, repeated cross-frame IDs, filtered IDs, and unconfirmed IDs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/trackers/io/mot.py	Validates unique ground-truth and tracker IDs per frame.
tests/io/test_mot.py	Tests duplicate rejection and valid exceptions.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.