Washed-out colors

[Anna-Szal]

Washed-out colors

[Anna-Szal]

Well, sort of.. The AHDRNet and all similar models merge 3 photos into one, where pixel values have range far greater than 256 (the HDR), which also means you cannot visualize it easily - your image viewer will probably just trim the values and you'll get a very dark image. So you need some kind of tone mapping (from high range to 256 values) - either do it manually in a Photoshop-like editor, or develop an algorithm (or another neural model). The authors here use a very primitive tone mapping operator, this one:
output = torch.log(1 + 5000 * output.cpu()) / torch.log(Variable(torch.from_numpy(np.array([1+5000])).float()))

Every other paper that used this operator with this dataset reported similar washed-out colored pictures, except for one - it had vivid colors, but the authors explicitly stated that they used image editor afterwards, so I guess it's safe to say that the authors of the AHDRNet did the same thing.

[Zasion]

Sincerely appreciate your answer : ) , I have solved this problem on MATLAB with tonemap function[tonemap(data,'AdjustLightness',[0,1],'AdjustSaturation',6)]. And I got a similar effect to the author's. My supervisor reminded me to read the paper carefully, I should have realized this problem earlier. Thank you again! Love from CN ❤️

[Zasion]

I encountered the same situation as you, did you solve it?

[Anna-Szal]

Huh, well that makes at least two of us, can't believe i missed this :D
I remember i used some openCV functions instead, but maybe this one is better :)