Problem about trajectory association in target assign process of tracking module

[yaruzz]

Thanks for your excellent job, I have some questions regarding the generation of 'matched_tracklet' information. The process of matching track results to ground truth trajectories using the Hungarian algorithm (implemented as ‘GNN_assignment’ in the code) implies that each pred_tracklet is matched with exactly one GT tracklet. However, when the tracks are fragmented, one GT track may correspond to multiple pred_tracklets. In such cases, the unmatched predicted tracklets are treated as negative samples for training in CRM. This approach seems unreasonable. Could you please clarify where my understanding went wrong?

[suojinhui]

So try to avoid the fragmentation of the trajectory.