PSNR Evaluation

[ThumbmasWalker]

Hi,

It was mentioned in issue #155 that to report PSNR on the DTU dataset, you do not use test/train splits and simply overfit to the training images.

If this is the case then why are the reported PSNR scores so much lower than those obtained by the model when overfitting to all training views? I'm trying to compare to your approach, but I feel unclear exactly how the PSNR scores reported in the paper were computed.

Thanks!