Why BEVDepthOCC gets worse results than BEVDetOCC ?

[JYounngS]

I changed the model type and img_view_transformer type in the config file, then I can only get mIoU of 27.84. Instead, if I always set the loss_depth to 0 while keeping the same config file, I can obtain mIoU of 31.9.

[JYounngS]

Has anyone tried to use the BEVDepth as your baseline?

[stduent-vehicle]

Have you tried it again? I am curious about it.

[zswzswzsw233]

I changed the model type and img_view_transformer type in the config file, then I can only get mIoU of 27.84. Instead, if I always set the loss_depth to 0 while keeping the same config file, I can obtain mIoU of 31.9.

应该是预训练权重的问题，没有提供bevdet-r50-depth-cbgs.pth这个权重