Awesome

Learning to See in the Dark in PyTorch.

This repo implements Learning to See in the Dark models presented in [1].

Test(Inference)

Usage:

python demo.py test <options>

In order to use pretrained models by authors of [1], which are trained using TensorFlow, specify a converted pretrained model file(see here) to --tf_weight_file.

If you use trained models by this repo, specify model files to --resume option.

• --tf_weight_file authors' TensorFlow models(see here)
• --resume trained models by this repo

Training

Usage:

python demo.py train <options>

Options

• cmd specifies train(training) or test(inference)
• --arch_type camera model type (default: Sony):
  • Sony for Sony camera(Sony alpha 7S II)
  • Fuji for Fujifilm camera(Fujifilm X-T2)
• --dataset_dir dataset directory(default: ./dataset/)
• --tf_weight_file weight file converted from authors' TensorFlow models(see here)(default: ``)
• --resume If set, checkpoint file used in train or test for PyTorch (default: ./checkpoint/Sony/)
• --checkpoint_dir checkpoints directory (default: ./checkpoint/Sony/)
• --result_dir directory where results are saved (default: ./result/Sony/)
• --train_img_list_file text file containing image file names for training
• --valid_img_list_file text file containing image file names for validation
• --test_img_list_file text file containing image file names for test
• --log_file log file (default: ./log/Sony/test.log)
• --gt_png If set, provided png images, converted from RAW format to png, are used as ground truth
• --batch_size batch size in training(default: 1)
• --valid_batch_size batch size in validation (default: 1)
• --test_batch_size batch size in test (default: 1)
• --patch_size patch size in training (default: None)
• --save_freq checkpoint save frequency (default: 1)
• --print_freq log print frequency (default: 1)
• --gpu GPU devide id (default: 0)
• -j or --workers number of data loading workers (default: 4)
• -c or --config configuration number(in which hyperparameters in training are set)

Pretrained models

The following weight files are models converted from authors' TensorFlow models listed in the first two rows, and models trained from scratch by codes in this repo(third row). I converted the authors TensorFlow checkpoint files to pickle format to load their weight values into PyTorch models.

Dataset

To download datasets by Sony and Fuji cameras, see authors' site.

PSNR/SSIM results

Table below shows PSNR and SSIM results using converted authors' TensorFlow models and a trained model by codes in this repo. Two types of test data are used:

• "all": all RAW test files are used. These files are listed in Sony_test_list.txt and Fuji_test_list.txt which are contained in dataset.
• "00 only": as written in here "00" means the number in the sequence/burst in the RAW file names like ./Sony/short/10003_00_0.04s.ARW.

Some Results

Sony

Camera: Sony a7S II

Info: ISO-2000 f/9 1/10s

Ground truth(RAW file: Sony/long/10106_00_30s.ARW)

The following images are inference and white balance results for RAW file: Sony/short/10106_00_0.1s.ARW

inference by the converted TensorFlow model

inference by the PyTorch model trained from scratch

white balance by rawpy(raw.postprocess(use_camera_wb=True, half_size=False, no_auto_bright=False, output_bps=16))

white balnce by Imagemagick convert command(e.g. convert arw:10106_00_0.1s.ARW 10106_00_0.1s.png)

Fuji

Camera: Fuji X-T2

Info: ISO-800 f/7.1 1/30s

Ground truth(RAW file: Fuji/long/10068_00_10s.RAF)

The following images are inference and white balance results for RAW file: Fuji/short/10068_00_0.033s.RAF

inference by the converted TensorFlow model

white balance by rawpy(raw.postprocess(use_camera_wb=True, half_size=False, no_auto_bright=False, output_bps=16))

white balnce by Imagemagick convert command(e.g. convert raf:10068_00_0.033s.RAF 10068_00_0.033s.png)

References

1. Chen, C., Chen, Q., Xu, J., & Koltun, V. (2018). Learning to See in the Dark. arXiv preprint arXiv:1805.01934.
   arXiv, github, Project Website