[Voxel-wise Refinement] How to re-sample N selected hard voxels features?

[mosjoker]

Great job. I have a question. Voxel-wise Refinement Module
Given N selected hard voxels，3D volume features dim：X' x Y' x Z' x D；N voxels selected in Pcoarse X x Y x Z x C.
How to re-sample their corresponding features from the fine-grained 3D volume features F? (coordinates？Can you explain in more detail?)

regard 3D volume features dim X' x Y' x Z' x D and Pcoarse dims X x Y x Z x C as identical(only scaling or cropping)? take the corresponding index or re-projection

Thank you

[songw-zju]

Thanks for your interest. In the actual implementation, we adopt the normalized coordinates of selected N voxels, and then use the mesh_grid function in Pytorch to sample features from the fine-grained voxel features.