search

dmarnerides / hdr-expandnet

Training and inference code for ExpandNet
Other
201 stars 40 forks source link

The domain of pre-train model #8

Closed vlee-harmonicinc closed 5 years ago

vlee-harmonicinc commented 5 years ago

According to the paper,

All the images contained linear RGB values

Is the pre-model (./weights.pth) trained in linear RGB domain? In this case, which domain is the input (image/video) of expand.py? It is assumed to be non-linear and processed to linear?

dmarnerides commented 5 years ago

Hi, The inputs are non-linear RGB LDR images Outputs are linear RGB HDR images. All are in [0, 1] range.

Thanks

vlee-harmonicinc commented 5 years ago

Thank you for your quick response. :) Is linear SDR - linear HDR also tested and verified that non-linear SDR input have better or similar result?

dmarnerides commented 5 years ago

The choice for non-linear SDR input is so that the method is fully end-to-end with no need to linearise first. However, linear SDR - linear HDR is not tested. I'm not sure how to do that though or if it's necessary, since we can't know the exact linearisation of the SDR (hence the gamma/tone mapping augmentation). Even applying a non-linearity can be considered tone mapping so it is encapsulated in the model.

Also, due to the nature of NNs, this linearisation of the SDR is not a problem, since they are highly expressive and non-linear functions. We can consider the linearisation as built-in the NN, without the need to explicitly define a separate step for it.