How is the Backbone for Extracting img Features Optimized ？

[liyih]

In the paper, you cat the pcd and img_feat into batch_fg_pcd_cams, and then voxelize batch_fg_pcd_cams. But voxelization will loss grad, so how does the parameter in the img_backbone be optimized?

Hope for your response!
Best