Dear organizers,
at the scoring script we can see:
"""
res_img = tone_map(np.load(res_img))
ref_img = tone_map(np.load(ref_img))
"""
therefore the submitted files must not have tone map applied, otherwise it will be applied twice, right? If that is the case, could you provide the function for the inverse tone mapping? Most models work in the tone mapping domain and I realized when I go back to the original domain (range [0,500]) my function generates artifacts. Also do you recommend clipping just in case? I share my function below
def itmo (x, c=0.25, clip=False):
out = - ((c*x) / (x-1))
return out
Thank you for your help.
Posted by: nanashi @ July 13, 2022, 10:20 a.m.Hi,
The evaluation at testing phase will be the same as validation phase. If your algorithms are working in tone-mapped domain, we provide the inverse tone-mapping function below for your reference.
def inv_tmap(x):
return x / (4 - 4*x)