fl82hope/N2D-transfer

N2D-transfer

Paper: Let There be Light: Improved Traffic Surveillance via Detail Preserving Night-to-Day Transfer

Datasets

include 1000 daytime images, 1000 nighttime images(night1, night2, night3, night4, each has 250 images.)

Requirements

Python 3.7.3
pyTorch 1.0.0

Train

1. For style augmentation, first aug all permutations first. Data is organized as following:

--Style
--Day

Run the code ./transfer.sh for style augmentation. Output will be stored in ./datasets/Cars_aug.

2. N2D transfer, run the code ./train.sh, where sample_path is the dataset path.

• Style mix framework:
• Night to day image translation vis results:

3. Detection evaluation model: faster_rcnn model

• N2D translation, detection results:
• after translation, detection vis results ploted on nighttime images:

4. Faster RCNN code is based on faster_rcnn code, please follow the project for compiling.

Bibtex

@article{fu2021auto,
      title={Let There be Light: Improved Traffic Surveillance via Detail Preserving Night-to-Day Transfer}, 
      author={Lan Fu and Hongkai Yu and Felix Juefei-Xu and Jinlong Li and Qing Guo and Song Wang},
      year={2021},
      journal={TCSVT}
}

@article{li2021domain,
  title={Domain adaptation from daytime to nighttime: A situation-sensitive vehicle detection and traffic flow parameter estimation framework},
  author={Jinlong Li and Zhigang Xu and Lan Fu and Xuesong Zhou and Hongkai Yu},
  journal={Transportation Research Part C: Emerging Technologies},
  year={2021}
}

Acknowledments

• WCT2
• Faster R-CNN