Awesome

RetinexRawMamba

Paper is available at arXiv

Clone this repository

git clone https://github.com/Cynicarlos/RetinexRawMamba.git
cd RetinexRawMamba

Environments Preparation

My cuda version: 11.7

conda create -n RetinexRawMamba python=3.9
conda activate RetinexRawMamba

pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

Download these two files from the following links and upload them to the server, and then install:
Note: find the corresponding veresion

1. causal_conv1d causal_conv1d-1.0.0+cu118torch2.0cxx11abiFALSE-cp39-cp39-linux_x86_64.whl
2. mamba_ssm mamba_ssm-1.0.1+cu118torch2.0cxx11abiFALSE-cp39-cp39-linux_x86_64.whl
   you can also download them easily here

pip install causal_conv1d-1.0.0+cu118torch2.0cxx11abiFALSE-cp39-cp39-linux_x86_64.whl  
pip install mamba_ssm-1.0.1+cu118torch2.0cxx11abiFALSE-cp39-cp39-linux_x86_64.whl

pip install -r requirements.txt

Dataset Preparation

Note that for SID Sony dataset, to be consistent with DNF, please use the Sony_test_list.txt we provide in the datasets folder to evaluate, and there are totally 562 images to be tested.
The directory for the datasets should be as following:

📁datasets/  
├─── 📁MCR/  
│    ├─── 📄MCR_test_list.txt  
│    ├─── 📄MCR_train_list.txt  
│    └─── 📁Mono_Colored_RAW_Paired_DATASET/  
│         ├─── 📁Color_RAW_Input/
│         │    ├─── 📄C00001_48mp_0x8_0x00ff.tif
│         │    └─── 📄...
│         └─── 📁RGB_GT/
│              ├─── 📄C00001_48mp_0x8_0x2fff.jpg
│              └─── 📄...
└─── 📁SID/  
     ├─── 📁Fuji/  
     │    ├─── 📄Fuji_test_list.txt  
     │    ├─── 📄Fuji_train_list.txt  
     │    ├─── 📄Fuji_val_list.txt  
     │    └─── 📁Fuji/  
     │         ├─── 📁Long/
     │         │    ├─── 📄00001_00_10s.RAF
     │         │    └─── 📄...
     │         └─── 📁Short/
     │              ├─── 📄00001_00_0.1s.RAF
     │              └─── 📄...
     └─── 📁Sony/  
          ├─── 📄Sony_test_list.txt  
          ├─── 📄Sony_train_list.txt  
          ├─── 📄Sony_val_list.txt  
          └─── 📁Sony/  
               ├─── 📁Long/
               │    ├─── 📄00001_00_10s.ARW
               │    └─── 📄...
               └─── 📁Short/
                    ├─── 📄00001_00_0.1s.ARW
                    └─── 📄...

Train from scratch

Before training and testing, please make sure the corresponding config file is correct, like the dataset dir, change it to your dataset path.

python train.py -cfg configs/sony.yaml

If you want to train on other dataset, just make sure you have the correct config file in the configs folder, and change the -cfg to your own config path.

Evaluate

Before evaluating our pretrained models, please download them by the following links and put them in the pretrained folder.

For MCR dataset:

python test_mcr.py

For SID dataset:
If your GPU memory is smaller than 40G, generally 24G, please use the following script so that you can test without OOM (out of memory).

python test_sony.py --merge_test

otherwise, ignore it. Note that the results may be a little bit smaller when merge testing than with whole image.

python test_sony.py

Citation

If there is any help for your research, please star this repository and if you want to follow this work, you can cite as follows:

@misc{chen2024retinexrawmambabridgingdemosaicingdenoising,
      title={Retinex-RAWMamba: Bridging Demosaicing and Denoising for Low-Light RAW Image Enhancement}, 
      author={Xianmin Chen and Peiliang Huang and Xiaoxu Feng and Dingwen Zhang and Longfei Han and Junwei Han},
      year={2024},
      eprint={2409.07040},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.07040}, 
}

Acknowledgement

The repository is refactored based on DNF, thanks to the author.