search

gabrieleilertsen / hdrcnn

HDR image reconstruction from a single exposure using deep CNNs
https://computergraphics.on.liu.se/hdrcnn/
BSD 3-Clause "New" or "Revised" License
502 stars 101 forks source link

Question about the format of the output? #47

Open HamidTohidypour opened 3 years ago

HamidTohidypour commented 3 years ago

Thanks for sharing your great work. Is the format of the generated HDR in the light domain or perceptual domain? is it in BT 2020? I used the network to get the HDR image in .exr format. The output values are very small for example the maximum value that I got was 38 for one of my images. Also, I tested the ground truth images provided on the hdrcnn's github page and the maximum value was 26 for one of the exr images and simialry for the other ground truth exr images the values were less than 100. Could you please let me know if the output of this network is an absolute HDR value and can be display as it is? If not and it is a normalized version of the actual values, how can I convert it to the absolute HDR value in the light domain or perceptual domain so that I can show it on an HDR display?

gabrieleilertsen commented 3 years ago

The HDR images are supposed to be linear, but not calibrated in absolute values. That is, to have an absolute calibration you would need to scale the images individually. For HDR display, I'm not sure what the best pre-processing would be, but you could for example scale an image to have a certain percentile scaled to the peak luminance of the display. So if 1 is mapped to the peak luminance, you could do I/percentile(I,95) to have the 5% brightest pixels clipped at the peak luminance. For the study we did in the paper, we fixed the luminance of the 90th percentile pixel value to 180 cd/m2.