BEVFormer
BEVFormer copied to clipboard
Questions about TemporalSelfAttention
I found when doing temporal alignment in TemporalSelfAttention, you rotate the previous bev feature and translate the reference points. Why not rotate and translate the previous bev feature or rotate and translate the reference points? Thanks.
Another question is, do you think the object queries instead of bev queries in DETR3D, also can make temporal alignment and fusion, will that work?
I have a question about the rotation of previous bev feature.
In line 152 in projects/mmdet3d_plugin/bevformer/modules/transformer.py:
tmp_prev_bev = rotate(tmp_prev_bev, rotation_angle, center=self.rotate_center)
Actually "rotation_angle" is just the yaw of prev_bev coord in global world. I don't know why rotating prev_bev by it.
Welcome to discuss. Thanks.