ByteTrack
ByteTrack
ByteTrack (2022) is a simple but effective MOT association algorithm that uses every detection bounding box, even low-confidence ones, instead of dropping them. Two-stage matching first associates high-score boxes with existing tracks, then uses low-score boxes to recover tracks that would otherwise be lost to occlusion or motion blur. ByteTrack…
Definition
ByteTrack (2022) is a simple but effective MOT association algorithm that uses every detection bounding box, even low-confidence ones, instead of dropping them. Two-stage matching first associates high-score boxes with existing tracks, then uses low-score boxes to recover tracks that would otherwise be lost to occlusion or motion blur. ByteTrack hits 80.3 MOTA on MOT17 with no learned re-ID network — pure IoU plus Kalman filter. FI Tech uses ByteTrack on aerial drone footage where re-ID features are unreliable from above and motion blur is constant; the low-score-recovery trick keeps trajectories continuous across rotor wash and atmospheric haze.
