Dear organizers,
where can we find the starter code or utils to load/save the images properly.
In the description it says the images have values in the range [0, 500], yet the maximum value in the ground-truth in the dataset is 111.92448.
After normalizing the images I have problems to re-scale the predictions back to the original npy range.
thank you.
Posted by: nanashi @ June 13, 2022, 4:21 p.m.Hi @,nanashi,
As described in "Get data" section, all evaluations are performed in tone-mapped domain. The tone mapping operation can be expressed as f(x) = x / (x+0.25), which is invertible.
This operation will cast all pixels in the range [0, 1]. You can refer to the tone_map function in evaluate.py in https://codalab.lisn.upsaclay.fr/my/datasets/download/b549bbbf-7c4e-4ada-bdfb-429523bac284.
Please do not normalize the data by dividing their maximum values. The range [0, 500] is for the whole dataset distribution, rather than each single image.
Hence, you have two options to deal with this issue: i) generate results in the original domain of range [0, 500] and save them as npy. ii) generate results in tone-mapped domain of range [0,1] and then rescale them back to the original domain.
Best,
Organizers