This code relies a lot on the projects GauGAN, SRGAN, BasicSR,pix2pix. Credit to these pytorch projects. Some preprocessing code for Tungsten scenes credits to Disney KPCN project.
experiments dir can be downloaded from googledrive
We provide two set of model weights. "./experiment/kjl/models/opt_XX.pth" is trained from KJL large indoor room datasets without seperating Specular and Diffuse components. You can finetune the weigths to your own dataset based on this pretrained weights.
"./experiments/tungsten_diffuse/models/opt_XX.pth" and "./experiments/tungsten_specular/models/opt_XX.pth" are finetuned on Tungsten scenes.
data dir can be downloaded from googledrive
This directory contain some samples from Tungsten scenes. It also provides the utility scripts to do data processing, for example, to process EXR files, Tungsten data pre-processing ect..
Large scale indoor dataset from Kujiale.com will be published soon.
script dir contains json files for training/testing
Attention!!! Change the settings in json files to include your own data path and project root Specifically, remember to change the "dataroot_NOISY" "dataroot_GT" "root" "val_root" to your paths
python train_diffuse.py -opt script/train/train_seperate_denoiser_diffuse.json
python train_specular.py -opt script/train/train_seperate_denoiser_specular.json
python test_diffuse.py -opt script/test/test_seperate_denoiser_diffuse.json
python test_specular.py -opt script/test/test_seperate_denoiser_specular.json