SysCV / qdtrack

Quasi-Dense Similarity Learning for Multiple Object Tracking, CVPR 2021 (Oral)
Apache License 2.0
384 stars 61 forks source link

About backdrops #110

Closed Leo63963 closed 2 years ago

Leo63963 commented 2 years ago

Hi! Really admire your great work, thanks for your sharing. I am quite confused about backdrops. The thing is, the backdrops are those unmatched detections, which are kept for one frame for potential matching. The embeddings and bboxes of backdrops are stored at self.backdrops, and "cat" with self.tracklets at self.memo. Therefore, the backdrops are regarded as one of tracklets to match with detections in the next frame, i.e., data association.

However, if matched by bi-softmax with a higher score, it still can not be matched successfully since the id of backdrops are all set to -1, which is violate the following code: https://github.com/SysCV/qdtrack/blob/b28af06c7fdb6ce99b967302c0c7e9a557d508bf/qdtrack/models/trackers/quasi_dense_embed_tracker.py#L188

Thus, the matching between backdrops and detections is unavailable. Is that true? Thanks.