about HDR data processing during training about fhdr

[mukulkhanna]

Hi @sunyclj,

Yes we aren't normalizing the ground truth (GT) HDR in the data loader. Its range is [0,inf] when it comes out of the data-loader. And that's okay because the GT HDR image doesn't go into the network. The GT HDR image is used for computing the loss when the network's prediction (from the LDR image) comes out.

We normalize the GT HDR image and the network's prediction both to [0,1] before we compute the loss between them (here).

This normalization is done using the mu-law tonemapping function (more info in the paper) implemented here.

[sunyclj]

@mukulkhanna ok,thank you. please, where is the code to calculate Q-score? I don't find.

[mukulkhanna]

We calculate the Q-score using the HDRVDP-2 (MATLAB) codebase. You can find more information about their project and source code here:

• http://hdrvdp.sourceforge.net/wiki/
• https://sourceforge.net/projects/hdrvdp/files/hdrvdp/