fix(eval): use MOT distractor classes in GT preprocessing for TrackEval parity

[RubenHaisma]

Description

The MOT ground-truth preprocessing parses the class column (mot.py) but never used it. Valid GT and distractor regions were both derived from the confidence flag alone:

valid_mask = frame_data.confidences > 0
...                       # valid GT
~valid_mask               # "distractor" == conf == 0

TrackEval's mot_challenge_2d_box.py keys off the class instead:

• GT kept for scoring = (conf != 0) AND (class == pedestrian, 1)
• distractor regions (tracker dets best-matching them are suppressed) = class ∈ {2, 7, 8, 12} — person_on_vehicle, static_person, distractor, reflection

Why it matters

On MOT17/MOT20 ground truth, distractor-class rows carry conf == 1. The old conf-only logic therefore:

• counted distractor-class boxes as ground truth → inflated FN / IDFN, and
• let tracker detections overlapping a distractor be scored as false positives (instead of being suppressed),

shifting MOTA / HOTA / IDF1 away from the official TrackEval numbers. The file documents TrackEval parity and _remove_distractor_matches cites the TrackEval MOT17 line range, so this is a parity gap rather than a design choice.

The existing _remove_distractor_matches machinery was already correct — it just needs the distractor mask to be class-based.

Scope / safety

This is invisible to the current integration tests, which cover only SportsMOT and DanceTrack (single pedestrian class, conf == 1). For that data the new and old masks are identical, so those results are unchanged. num_tracker_ids is intentionally left as-is (built before suppression); a never-matched id is metric-neutral.

Tests

tests/io/test_mot.py (hermetic, no downloads):

• test_distractor_class_excluded_and_matching_tracker_removed — a distractor-class GT (conf==1) and a conf==0 pedestrian are excluded from GT, and a tracker det overlapping the distractor is suppressed. Fails on the pre-fix code (num_gt_dets == 2), passes after.
• test_single_class_sequence_unaffected — SportsMOT/DanceTrack-style data passes through unchanged.

[Copilot]

Pull request overview

This PR updates MOT-format ground-truth preprocessing to match TrackEval’s MOTChallenge behavior by using the GT class column to decide which GT rows are scored and which regions act as distractors for suppression, improving metric parity (e.g., MOTA/HOTA/IDF1) on MOT-style datasets.

Changes:

• Add class-based ground-truth masks to keep only pedestrian-class, non-ignored GT rows for scoring.
• Change distractor handling to use MOT distractor classes for suppressing tracker detections matched to distractor regions.
• Add unit tests to cover distractor-class GT with conf==1 and ensure single-class sequences remain unchanged.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/trackers/io/mot.py	Introduces helper masks for valid GT vs distractor GT and wires them into sequence preparation for TrackEval parity.
tests/io/test_mot.py	Adds hermetic tests validating distractor-class exclusion/suppression and unchanged behavior for single-class data.