squeeze cls

[frkmac3]

SparseDrive/projects/mmdet3d_plugin/models/detection3d/decoder.py

Line 49 in ddb706b

cls_scores, cls_ids = cls_scores.max(dim=-1)

I noticed that when instance_id is not None, you do this step first，can you tell me why

[swc-17]

In line 54, cls_score is flattened from (bs, num_pred, num_cls) to (bs, num_pred * num_cls), then top k predictions are selected through higher score. In this case, if one instance predicts high score for all classes, the corresponding box will be selected several times, it's ok when performing detection task only, however, if we want to perform tracking simultaneously, we want this instance only be selected once and maintains the unique instance id, so squeeze cls is needed.