BEV channel-compressed from 256 to 4 #4
Description
Considering the future prediction task requires low-dimension inputs, we further compress the fused BEV feature into a low-dimensional(C′ = 4) BEV feature.
Hi, I noticed that the BEV feature channels after LiDAR-camera fusion are 256, and then they are compressed to be compatible with the diffusion model. Could you please elaborate on how this compression is performed? Given that the channels are reduced by such a large factor, have you evaluated the potential loss in reconstruction performance?