Closed peiyaoooo closed 2 years ago
@peiyaoooo for the first question, this operation aims to normalize LDR valuea to the range [-1, 1], which is the common method for training.
And we use OpenEXR format to represent the HDR image, which is compatible with opencv imread function (the arg '-1' can handle this).
Oh thank you very much! And I just found that I made a silly error. It is "LDR = LDR **2 -1" in my head and I couldn't understand this comand. It's my fault. So sorry about this. And thank you for you reply again.
Hi, thank you for this amazing work! And I have some questions about the LDR and HDR pre-processing to ensure scale-invariant.
In https://github.com/LWT3437/LANet/blob/d45c86406208f81105e3960889b28ed9b50f8ad6/LANet/src/utils.py#L46, The linear LDR image is squared and then subtracted by 1. I could't get the aim of this operation.
Besides, in the loading procedure (https://github.com/LWT3437/LANet/blob/d45c86406208f81105e3960889b28ed9b50f8ad6/LANet/src/utils.py#L21), the LDR and HDR images are all loaded via cv2. But the LDR and HDR image are in different format right? like LDR in 8 bit and HDR in 14/16 bit.
Hope for your reply and thank you again.