CyclingTrack

CyclingTrack is a benchmark dataset for multi-object tracking in aerial/broadcast footage of professional road cycling. Sequences feature dense pelotons with small rider scale, moving cameras, rapid viewpoint changes, varying illumination, and frequent occlusions—conditions underrepresented in ground-view datasets. Annotations follow the MOTChallenge format (gt.txt) and were produced semi-automatically in CVAT: baseline tracker outputs were corrected to remove false positives/negatives, refine bounding boxes, and fix identity switches. The release provides per-frame images and ground-truth annotations with train and test splits to enable reproducible evaluation of tracking-by-detection methods under broadcast constraints. Recommended metrics: HOTA, IDF1, and MOTA (e.g., via TrackEval).