Closed yj772881654 closed 1 week ago
During training, the false positive track queries are some noisy instances that not propagated from the matched positive predictions of the last frame. They are randomly sampled from the negative predictions from the last frame. The goal is to mimic the wrong predictions of the last frame, which naturally happens at test time. However, we actually turn this off (see the config) following MOTRv2's practices (See Sec4.2 of their paper).
Thank you for your response; it resolved my doubts. I have another question. Where are the Deformable Self Attention and Perspective-to-BEV Cross Attention in the BEV Module implemented? I don't seem to see the relevant content in MapTracker.py.
Thank you for your response; it resolved my doubts. I have another question. Where are the Deformable Self Attention and Perspective-to-BEV Cross Attention in the BEV Module implemented? I don't seem to see the relevant content in MapTracker.py.
Those are in the bevformer folder
I have another question about the function prepare_track_queries_and_targets. What is the role of track_queries_fal_pos_mask and false positive queries, and how do these differ from the normal track queries?