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

Measure the reconstruction and the ground truth using HDR-VDP-2.2 #35

Closed PK15946 closed 4 years ago

PK15946 commented 4 years ago

Hi @gabrieleilertsen Can those reconstructed exr files directly compare with the ground truth exr files using HDR-VDP?
I am trying to directly measure the reconstruction as followed

ref = hdrread('the exr ground truth in the provided testset')
tst = hdrread('your reconstruction exr file')
hdrvdp(tst,ref,'rgb-bt.709',30)

but the average Q score is much greater than it reported in ExpandNet(59 vs 41), I am not sure the result is right consider such a big difference. Or, should I calibrate those exr file in absolute luminance BEFORE measuring them? like pfsin relative.exr | pfstmo_mantiuk08 -e 1 -d l=4000:b=0.01:a=0 | pfsdisplayfunction -d l=4000:b=0.01:a=0 -l | pfsout absolute.exr

gabrieleilertsen commented 4 years ago

The result is very much depending on the calibration and intended setup. HDR-VDP expects calibrated pixels as input, so the images need to be scaled to approximately correspond to absolute luminance values. The intended viewing setup also influences the result, where e.g. more pixels per visual angle (more close up view) means lower score.

It could be difficult to compare results to previous evaluations without knowing the setup, but you can still use HDR-VDP for evaluating the results of different algorithms given the same evaluation setup.

PK15946 commented 4 years ago

Got it! Thanks a lot!