Subset: the 7 training sequences: MOT17-02, -04, -05, -09, -10, -11, -13 (5,330 frames, 640x480 to 1920x1080, 14-30 fps).
Source of the copy: the public Hugging Face mirror Lekim89/MOT17 (the official host was unreachable from the
development environment); only the -FRCNN copy of each sequence is fetched since the three detector variants share images.
Annotations used: pedestrian class, consider flag = 1, with visibility; no distractor handling.
Licence: CC BY-NC-SA 3.0 (MOTChallenge). The repository's code is MIT; the clips, ground-truth overlays and detections in
web/public/data/ and the GIFs in docs/assets/ are derived from MOT17 and carry its licence.
Contents of concern: real people filmed in public places; faces are visible. The project produces no identities.
Splits: a temporal split (trailing 25% of MOT17-02) for tuning only; evaluation on whole sequences.