Dear Edu,
We are running our program right now, since we want to validate our evaluation way is correct, I am a little bit cofusing about the no processing result, may I know how to get this? basicly just using the medial exposure one as output?
Many thanks,
Bozhi
Posted by: luciferzhi @ Feb. 16, 2022, 12:50 p.m.Dear Bozhi,
In principle a "No processing" result would be using the reference frame as output (this is generally the medium frame, which is spatially and temporally aligned to the ground-truth image).
Best,
Edu.
Dear Edu,
We have tried to calulate the PSNR with medium exposure image with groundtruth, however there is always something wrong. We normailized the input SDR image and the groundtruth HDR image to [0,1] by their maximum value, then caculate the PSNR, but we always get a small PNSR compared to u-PSNR, like psnr:14db mu-psnr:25db. Since the reference SDR image is 8bits with a range [0,1], however the groundtruth HDR image is 16bits with a range [0,10], would you mind to point out which way we are wrong about this?
Many thanks,
Bozhi
Posted by: luciferzhi @ March 18, 2022, 1:29 p.m.Dear Bozhi,
Please use the metrics implementations provided within the challenge (i.e. metrics.py, data_io.py) in case you are not already using them. In order to compute the two PSNR metrics, you can just run:
# This snipped assumes necessary imports from metrics.py, data_io.py and numpy
normalized_psnr(ref_hdr_image, res_hdr_image, np.max(ref_hdr_image))
psnr_tanh_norm_mu_tonemap(ref_hdr_image, res_hdr_image, percentile=99, gamma=2.24)
where ref_hdr_image is the ground-truth image as read by the imread_uint16_png( ) function and res_hdr_image is your estimated image.
Best regards,
Edu.