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
508 stars 101 forks source link

preprocess of training data #28

Closed durgob closed 5 years ago

durgob commented 5 years ago

Hi, Thanks for your impressive job!

I have downloaded some training data with your provided dataset list(Thanks a again, that save me a lot of time), some are .exr files, and the other are .hdr files.

Accidentally I found that the max value of these files are different: some seem like they have been normalized(value in [0, 1]), and some haven't(value in [0, hundreds]).

But there is not normalize for these files in virtualcamera because they all are in 32FC3 format.

So, should I normalize them into [0, 1] before processed by virtualcamera?

gabrieleilertsen commented 5 years ago

Hi. The normalization is taken care of by the virtualcamera script. This is done on lines 237-274, which applies a random scaling so that a certain percentile of the pixels will be considered saturated and clipped when thresholding at 1. So it does not matter how your input images are scaled.

durgob commented 5 years ago

@gabrieleilertsen Thanks for your reply, but I am still a bit confused.

Assuming clip is set as 20%. In normalized case, let's say value of pixels in [0, 0.8] occupy 80% of the number of pixels, then sc = 0.8, and the scaled host_roi in range[0, 1.25] If I am not mistaken.

When Not normalized, let's say the value of pixels is between 0 and 100. And value in [0, 0.8] occupy 80%, then the scaled host_roi in range[0, 125].

So there is a big difference in term of max value(1.25 vs 125), will it impede the training process?

gabrieleilertsen commented 5 years ago

Yes, that is correct. Both images will have the same number of saturated pixels, but different maximum values. The max value can vary quite a lot between images, but this is part of the LDR to HDR problem. Also, using a loss function in the log domain makes easier to handle extremely large values so that they do not influence too much.

durgob commented 5 years ago

Got it! @gabrieleilertsen Thanks a lot!