PCGAN_Model

Prerequisites

Linux
NVIDIA GPU + CUDA CuDNN
PyTorch >=1.1.0

Train

You can run our training code in gvien examples of dataset folder by following instruction.

python train.py  --dataroot=./dataset --batch_size=1

Models are saved to ./checkpoints/.

See opt in files(base_options.py and train_options.py) for additional training options.

Test

You can test our pretrained model in gvien examples of dataset folder by following instruction.

python test.py --dataroot=./dataset --epoch=200 --data_type=short

'--epoch=200' means loading the 200th epoch saved model.

'--data_type=short or long', 'short' means testing imges of short exposure time, while 'long' means testing imges of long exposure time (saturation images).

The pretrained model can be downloaded at Baidu Netdisk, password: tr2n.

Testing results are saved in ./result/， and each result inculdes 8 images, I_A, I_M, I_G, I_comp, I_GT, I_SAT, M(I_G), M(I_GT), Where the mean of these tokens are in our paper.

See opt in files (base_options.py and test_options.py) for additional testing options.

Dataset

Several samples of our dataset are in ./dataset. Each sample incules 3 images, I_SAT, I_GT, I_M. The size of training image is 350 x 350, but they are randomly crop to 256 x 256 in training process. Testing image is 256 x 256.

Acknowledgments