According to the documentation, it seems like kaolin.ops.spc.to_dense() returns a tensor of size [bs, feature_dim, 2^l, 2^l, 2^l]. However, this seems weird as a feature grid should be of size [bs, feature_dim, 2^l+1, 2^l+1, 2^l+1] since we are effectively dealing with the dual of voxels (i.e. corners). Is this because this function returns the interpolated voxel centers? If so, it would be quite useful to also have the option to return the dense feature grids "as is". In particular, this would facilitate the implementation of a progressive upsampling approach à la NSVF and similar works.
According to the documentation, it seems like
kaolin.ops.spc.to_dense()returns a tensor of size[bs, feature_dim, 2^l, 2^l, 2^l]. However, this seems weird as a feature grid should be of size[bs, feature_dim, 2^l+1, 2^l+1, 2^l+1]since we are effectively dealing with the dual of voxels (i.e. corners). Is this because this function returns the interpolated voxel centers? If so, it would be quite useful to also have the option to return the dense feature grids "as is". In particular, this would facilitate the implementation of a progressive upsampling approach à la NSVF and similar works.