Skip to content

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision

Notifications You must be signed in to change notification settings

bupt-ai-cz/LLVIP

Repository files navigation

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision visitors

Project | Arxiv | Kaggle | PWC | Tweet

News

  • ⚡(2024-1-8): The pre-trained pix2pixGAN on LLVIP is released Here.
  • ⚡(2023-2-21): The annotations of a small part of images have been corrected and updated, including the annotation of some missing pedestrians, and the optimization of some imprecise annotations. The updated dataset is now available from the homepage or here. If you need the previous version of the annotations, please refer to here.
  • ⚡(2022-5-24): We provide a toolbox for various format conversions (xml to yolov5, xml to yolov3, xml to coco)
  • ⚡(2022-3-27): We released some raw data (unregistered image pairs and videos) for further research including image registration. Please visit homepage to get the update. (2022-3-28 We have updated the link of Baidu Yun of LLVIP raw data, the data downloaded from the new link supports decompression under windows and macos. The original link only support windows.)
  • ⚡(2021-12-25): We released a Kaggle Community Competition "Find Person in the Dark!" based on part of LLVIP dataset. Welcome playing and having fun! Attention: only the visible-image data we uploaded in Kaggle platform is allowed to use (the infrared images in LLVIP or other external data are forbidden)
  • ⚡(2021-11-24): Pedestrian detection models were released
  • ⚡(2021-09-01): We have released the dataset, please visit homepage or here to get the dataset. (Note that we removed some low-quality images from the original dataset, and for this version there are 30976 images.)

figure1-LR


Dataset Downloading:


Citation

If you use this data for your research, please cite our paper LLVIP: A Visible-infrared Paired Dataset for Low-light Vision:

@inproceedings{jia2021llvip,
  title={LLVIP: A visible-infrared paired dataset for low-light vision},
  author={Jia, Xinyu and Zhu, Chuang and Li, Minzhen and Tang, Wenqi and Zhou, Wenli},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3496--3504},
  year={2021}
}

or

@misc{https://doi.org/10.48550/arxiv.2108.10831,
  doi = {10.48550/ARXIV.2108.10831}, 
  url = {https://arxiv.org/abs/2108.10831},
  author = {Jia, Xinyu and Zhu, Chuang and Li, Minzhen and Tang, Wenqi and Liu, Shengjie and Zhou, Wenli}, 
  keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {LLVIP: A Visible-infrared Paired Dataset for Low-light Vision},
  publisher = {arXiv},
  year = {2021},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

Image Fusion

Baselines

FusionGAN

Preparation

  • Install requirements
    git clone https://github.com/bupt-ai-cz/LLVIP.git
    cd LLVIP/FusionGAN
    # Create your virtual environment using anaconda
    conda create -n FusionGAN python=3.7
    conda activate FusionGAN
    
    conda install matplotlib scipy==1.2.1 tensorflow-gpu==1.14.0 
    pip install opencv-python
    sudo apt install libgl1-mesa-glx
  • File structure
    FusionGAN
    ├── ...
    ├── Test_LLVIP_ir
    |   ├── 190001.jpg
    |   ├── 190002.jpg
    |   └── ...
    ├── Test_LLVIP_vi
    |   ├── 190001.jpg
    |   ├── 190002.jpg
    |   └── ...
    ├── Train_LLVIP_ir
    |   ├── 010001.jpg
    |   ├── 010002.jpg
    |   └── ...
    └── Train_LLVIP_vi
        ├── 010001.jpg
        ├── 010002.jpg
        └── ...
    

Train

python main.py --epoch 10 --batch_size 32

See more training options in main.py.

Test

python test_one_image.py

Remember to put pretrained model in your checkpoint folder and change corresponding model name in test_one_image.py. To acquire complete LLVIP dataset, please visit https://bupt-ai-cz.github.io/LLVIP/.

Densefuse

Preparation

  • Install requirements
    git clone https://github.com/bupt-ai-cz/LLVIP
    cd LLVIP/imagefusion_densefuse
    
    # Create your virtual environment using anaconda
    conda create -n Densefuse python=3.7
    conda activate Densefuse
    
    conda install scikit-image scipy==1.2.1 tensorflow-gpu==1.14.0
  • File structure
    imagefusion_densefuse
    ├── ...
    ├──datasets
    |  ├──010001_ir.jpg
    |  ├──010001_vi.jpg
    |  └── ...
    ├──test
    |  ├──190001_ir.jpg
    |  ├──190001_vi.jpg
    |  └── ...
    └──LLVIP
       ├── infrared
       |   ├──train
       |   |  ├── 010001.jpg
       |   |  ├── 010002.jpg
       |   |  └── ...
       |   └──test
       |      ├── 190001.jpg
       |      ├── 190002.jpg
       |      └── ...
       └── visible
           ├──train
           |   ├── 010001.jpg
           |   ├── 010002.jpg
           |   └── ...
           └── test
               ├── 190001.jpg
               ├── 190002.jpg
               └── ...
    

Train & Test

python main.py 

Check and modify training/testing options in main.py. Before training/testing, you need to rename the images in LLVIP dataset and put them in the designated folder. We have provided a script named rename.py to rename the images and save them in the datasets or test folder. Checkpoints are saved in ./models/densefuse_gray/. To acquire complete LLVIP dataset, please visit https://bupt-ai-cz.github.io/LLVIP/.

IFCNN

Please visit https://github.com/uzeful/IFCNN.

Pedestrian Detection

Baselines

Yolov5

Preparation

Linux and Python>=3.6.0

  • Install requirements

    git clone https://github.com/bupt-ai-cz/LLVIP.git
    cd LLVIP/yolov5
    pip install -r requirements.txt
  • File structure

    The training set of LLVIP is used for training the yolov5 model and the testing set of LLVIP is used for the validation of the yolov5 model.

    yolov5
    ├── ...
    └──LLVIP
       ├── labels
       |   ├──train
       |   |  ├── 010001.txt
       |   |  ├── 010002.txt
       |   |  └── ...
       |   └──val
       |      ├── 190001.txt
       |      ├── 190002.txt
       |      └── ...
       └── images
           ├──train
           |   ├── 010001.jpg
           |   ├── 010002.jpg
           |   └── ...
           └── val
               ├── 190001.jpg
               ├── 190002.jpg
               └── ...
    

    We provide a toolbox for converting annotation files to txt files in yolov5 format.

Train

python train.py --img 1280 --batch 8 --epochs 200 --data LLVIP.yaml --weights yolov5l.pt --name LLVIP_export

See more training options in train.py. The pretrained model yolov5l.pt can be downloaded from here. The trained model will be saved in ./runs/train/LLVIP_export/weights folder.

Test

python val.py --data --img 1280 --weights last.pt --data LLVIP.yaml

Remember to put the trained model in the same folder as val.py.

Our trained model can be downloaded from here: Google-Drive-Yolov5-model or BaiduYun-Yolov5-model (code: qepr)

Results

We retrained and tested Yolov5l and Yolov3 on the updated dataset (30976 images).

Where AP means the average of AP at IoU threshold of 0.5 to 0.95, with an interval of 0.05.

The figure above shows the change of AP under different IoU thresholds. When the IoU threshold is higher than 0.7, the AP value drops rapidly. Besides, the infrared image highlights pedestrains and achieves a better effect than the visible image in the detection task, which not only proves the necessity of infrared images but also indicates that the performance of visible-image pedestrian detection algorithm is not good enough under low-light conditions.

We also calculated log average miss rate based on the test results and drew the miss rate-FPPI curve.

Image-to-Image Translation

Baseline

pix2pixGAN

Preparation

  • Install requirements
    cd pix2pixGAN
    pip install -r requirements.txt
  • Prepare dataset
  • File structure
    pix2pixGAN
    ├── ...
    └──datasets
       ├── ...
       └──LLVIP
          ├── train
          |   ├── 010001.jpg
          |   ├── 010002.jpg
          |   ├── 010003.jpg
          |   └── ...
          └── test
              ├── 190001.jpg
              ├── 190002.jpg
              ├── 190003.jpg
              └── ...
    

Train

python train.py --dataroot ./datasets/LLVIP --name LLVIP --model pix2pix --direction AtoB --batch_size 8 --preprocess scale_width_and_crop --load_size 320 --crop_size 256 --gpu_ids 0 --n_epochs 100 --n_epochs_decay 100

Test

python test.py --dataroot ./datasets/LLVIP --name LLVIP --model pix2pix --direction AtoB --gpu_ids 0 --preprocess scale_width_and_crop --load_size 320 --crop_size 256

See ./pix2pixGAN/options for more train and test options.


Results

We retrained and tested pix2pixGAN on the updated dataset(30976 images). The structure of generator is unet256, and the structure of discriminator is the basic PatchGAN as default.

License

This LLVIP Dataset is made freely available to academic and non-academic entities for non-commercial purposes such as academic research, teaching, scientific publications, or personal experimentation. Permission is granted to use the data given that you agree to our license terms.

Call For Contributions

Welcome to point out errors in data annotation. If you want to modify the label, please refer to the annotation tutorial, and email us the corrected label file.

More annotation forms are also welcome (such as segmentation), please contact us.

Acknowledgments

Thanks XueZ-phd for his contribution to LLVIP dataset. He corrected the imperfect annotations in the dataset.

Contact

email: shengjie.Liu@bupt.edu.cn, czhu@bupt.edu.cn, jiaxinyujxy@qq.com, tangwenqi@bupt.edu.cn