/cchen156/Learning-to-See-in-the-Dark

cchen156/Learning-to-See-in-the-Dark


This is a Tensorflow implementation of Learning to See in the Dark in CVPR 2018, by Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun.

Project Website
Paper

teaser

This code includes the default model for training and testing on the See-in-the-Dark (SID) dataset.

Demo Video

https://youtu.be/qWKUFK7MWvg

Setup

Requirement

Required python (version 2.7) libraries: Tensorflow (>=1.1) + Scipy + Numpy + Rawpy.

Tested in Ubuntu + Intel i7 CPU + Nvidia Titan X (Pascal) with Cuda (>=8.0) and CuDNN (>=5.0). CPU mode should also work with minor changes but not tested.

Dataset

Update Aug, 2018: We found some misalignment with the ground-truth for image 10034, 10045, 10172. Please remove those images for quantitative results, but they still can be used for qualitative evaluations.

You can download it directly from Google drive for the Sony (25 GB) and Fuji (52 GB) sets.

There is download limit by Google drive in a fixed period of time. If you cannot download because of this, try these links: Sony (25 GB) and Fuji (52 GB).

New: we provide file parts in Baidu Drive now. After you download all the parts, you can combine them together by running: “cat SonyPart* > Sony.zip” and “cat FujiPart* > Fuji.zip”.

The file lists are provided. In each row, there are a short-exposed image path, the corresponding long-exposed image path, camera ISO and F number. Note that multiple short-exposed images may correspond to the same long-exposed image.

The file name contains the image information. For example, in “10019_00_0.033s.RAF”, the first digit “1” means it is from the test set (“0” for training set and “2” for validation set); “0019” is the image ID; the following “00” is the number in the sequence/burst; “0.033s” is the exposure time 1/30 seconds.

Testing

  1. Clone this repository.
  2. Download the pretrained models by running
python download_models.py
  1. Run “python test_Sony.py”. This will generate results on the Sony test set.
  2. Run “python test_Fuji.py”. This will generate results on the Fuji test set.

By default, the code takes the data in the “./dataset/Sony/” folder and “./dataset/Fuji/”. If you save the dataset in other folders, please change the “input_dir” and “gt_dir” at the beginning of the code.

Training new models

  1. To train the Sony model, run “python train_Sony.py”. The result and model will be save in “result_Sony” folder by default.
  2. To train the Fuji model, run “python train_Fuji.py”. The result and model will be save in “result_Fuji” folder by default.

By default, the code takes the data in the “./dataset/Sony/” folder and “./dataset/Fuji/”. If you save the dataset in other folders, please change the “input_dir” and “gt_dir” at the beginning of the code.

Loading the raw data and proccesing by Rawpy takes significant more time than the backpropagation. By default, the code will load all the groundtruth data processed by Rawpy into memory without 8-bit or 16-bit quantization. This requires at least 64 GB RAM for training the Sony model and 128 GB RAM for the Fuji model. If you need to train it on a machine with less RAM, you may need to revise the code and use the groundtruth data on the disk. We provide the 16-bit groundtruth images processed by Rawpy: Sony (12 GB) and Fuji (22 GB).

Citation

If you use our code and dataset for research, please cite our paper:

Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun, “Learning to See in the Dark”, in CVPR, 2018.

License

MIT License.

FAQ

  1. Can I test my own data using the provided model?

The proposed method is designed for sensor raw data. The pretrained model probably not work for data from another camera sensor. We do not have support for other camera data. It also does not work for images after camera ISP, i.e., the JPG or PNG data.

  1. Will this be in any product?

This is a research project and a prototype to prove a concept.

  1. How can I train the model using my own raw data?

Generally, you just need to subtract the right black level and pack the data in the same way of Sony/Fuji data. If using rawpy, you need to read the black level instead of using 512 in the provided code. The data range may also differ if it is not 14 bits. You need to normalize it to [0,1] for the network input.

  1. Why the results are all black?

It is often because the pre-trained model not downloaded properly. After downloading, you should get 4 checkpoint related files for the model.

Questions

If you have additional questions after reading the FAQ, please email to cchen156@illinois.edu.

Original Source