Compare Quality Score about 32-bit hdr image

[Skyler]

Hello, Rafal

 

I'm using HDR-VDP-3 to calculate the Q-Score.

The input images should normalize to [0, 1]. But I have multiple test images to compare the reference image. All of them are 32-bit hdr images, and I directly divide the range's maximum. The lower value will not fit the bit depth and become zero.

How do I use the correct method to normalize the test & reference images?

Thank you so much.

Best regards

Skyler.

[Rafal]

I have already replied by email - now I am reposting for the benefit of others.

HDR-VDP requires that the pixel values are provided in the absolute photometric units:

https://hdrvdp.sourceforge.net/wiki/index.php/Frequently_Asked_Questions

The simplest approach is to pass:

exposure = 4000/max(img_reference(:));

hdrvdp3( ..., img_test * exposure, img_reference * exposure, ... )

If the metric is to be used to evaluate single-image-HDR reconstruction methods, I recommend using the camera response curve correction from: 

https://www.cl.cam.ac.uk/research/rainbow/projects/sihdr_benchmark/

In that case, I suggest that you apply the exposure compensation above only on the reference image (in the linear colour space) and then use:

https://github.com/gfxdisp/pu21/blob/main/matlab/crf_correction.m

to align the test image.

Rafal