Task Formulation

Different from prior fully-/semi-supervised low-light detection paradigms, we work in a novel setting called Zero-Shot Day-Night Domain Adaptation (ZSDA). During training, we only have access to the well-lit images. Then, the trained model is directly evaluated on low-light images.

Alt text

Overview

During training, our pipeline takes only well-lit images as input, and learns low-light object detection. Given the synthesized well-lit/low-light image pairs and their corresponding image decomposition pseudo GT, our DAI-Net learns to predict reflectance maps through a reflectance decoder, therefore encoding illumination-invariant information into its base detector. Moreover, we further reinforce the reflectance representation with a mutual feature alignment loss and an interchange-redecomposition-coherence procedure.

Branching

Results

Experimental Setting: Source dataset (train) -> Target dataset (target)

Face Detection

Wider Face / COCO -> Dark Face
Alt text

Object Detection

COCO -> ExDark

alt

Visualization

Detection results on Dark Face and ExDark

alt

Visualization of original image, pseudo ground truth, and predicted reflectance

alt

Citation

@inproceedings{du2024boosting,
  title={Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation},
  author={Du, Zhipeng and Shi, Miaojing and Deng, Jiankang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12666--12676},
  year={2024}
}

or

@article{du2023boosting,
  title={Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation},
  author={Du, Zhipeng and Shi, Miaojing and Deng, Jiankang},
  journal={arXiv preprint arXiv:2312.01220},
  year={2023}
}