Training Linearization Net

[aasharma90]

Hi @alex04072000

I've a question about training Linearization Net. I see that -
First, you apply a randomly selected CRF crf (and its corresponding inverse CRF is invcrf) to convert an HDR image into a LDR image

Second, after some processing of the LDR image, you predict the inverse CRF from it (pred_invcrf) which is matched with the true inverse CRF (invcrf)

I could be wrong, but I think an HDR image is not necessarily similar to a RAW image, which basically means that it could already have some non-linear CRF of its own. Now, when you further apply crf, the non-linear CRF that would occur in the LDR image would also factor in the already existing non-linear CRF in the HDR image. This means that the "correct" inverse CRF expected from the LDR image cannot be invcrf. Could you please let me know if I'm understanding anything incorrectly?

Ofcourse, if in your dataset, you have made all HDR composites using RAW images, then it's completely fine. Thanks!