EBSNetMEFNet

This repo includes the source code of the paper: Learning a Reinforced Agent for Flexible Exposure Bracketing Selection (CVPR 2020) by Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, Jimmy Ren.

Quick Test

The code is tested on 64 bit Linux (Ubuntu 14.04 LTS) and based on Pytorch 0.4.1 with Python 2.7.

1. Clone this github repo

    git clone https://github.com/wzhouxiff/EBSNetMEFNet.git
    cd EBSNetMEFNet
   

2. Download models and testset from Baidu Drive (extraction code: jqfp) or Google Drive. Models are in folder checkpoints which testset is in folder testset.

3. Update scripts/test.sh with your path.

    usage: test.py [-h] [--data-type] [--results PATH] [--score-path PATH]
                    DIR DIR
   
    PyTorch EBSNetMEFNet
   
    positional arguments:
    DIR                       path to testset
    DIR                       path to models
   
    optional arguments:
    -h, --help                show this help message and exit
    --data-type               'night' or 'day'
    --results                 path to save results
    --score-path              path to save psnr and ssim
   

4. Run scripts/test.sh.

    sh scripts/test.sh
   

EBSNet v.s. MEFNet

EBSNet - Exposure Bracketing Selection Network: Used for exposure bracketing selection by analyzing both the illumination and semantic information of an auto-exposure preview image and Learnt via RL which rewarded by MEFNet.

MEFNet - Multi-Exposure Fusion Network: Used for fusing the selected exposure bracketing predicted by EBSNet.

These two networks can be trained jointly.

Dataset

• x: AE image
• z₀ ~ z₉: exposure sequence
• zz_H: generated HDR image
• testset - extraction code: jqfp

Results

Citation

@inproceedings{Wang2020Learning,
    title={Learning a Reinforced Agent for Flexible Exposure Bracketing Selection},
    author={Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, Jimmy Ren},
    booktitle={CVPR},
    year={2020},
}

Contributing

For any questions, feel free to open an issue or contact us (zhouzi1212@gmail.com)