PJLab-ADG / DetZero

[ICCV 2023] DetZero: Rethinking Offboard 3D Object Detection with Long-term Sequential Point Clouds
https://superkoma.github.io/detzero-page/
Apache License 2.0
334 stars 32 forks source link

Problem about trajectory association in target assign process of tracking module #56

Open yaruzz opened 6 months ago

yaruzz commented 6 months ago

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?