From ICCV paper...
Input Image
Reconstructed
Volume