A Benchmark for face hallucination in low-light scenarios

Face hallucination in low-light environments is an extremely challenging task due to the significant loss of facial structure and facial texture information. We construct a low-light face hallucination dataset DARKFFHQ-4000 based on the face dataset FFHQ as a benchmark for low-light face hallucination task. This dataset will provide researchers with a standardized and uniform benchmark for training and evaluation.

We select 4000 face images from the FFHQ dataset and downsample the HR face images using the Bicubic degradation model with scale factors of 4 and 8, resulting in 4000 face images with 64 $\times$ 64 pixels and 4000 face images with 32 $\times$ 32 pixels. These two sets of images are used as the LR face image. After that, we synthesize low-light LR face images based on the LR face images. We randomly select parameters $\alpha$, $\beta$, and $\gamma$ that are used to control contrast, brightness, and gamma correction in three well-designed ranges during the synthesis of each image. The process of generating low-light LR face images can be formulated as ${LR}_{low}(x,y) = \left(\frac{\alpha \times LR{(x,y)} + \beta}{255}\right)^{\gamma} \times 255$. More details are provided in Learning to Hallucinate Face in the Dark. Some samples are shown below:

The dataset can be downloaded from Google Drive, which contains HR, LR (scale factors are 4 and 8), and low-light LR face images.

If you find this dataset helpful in your research or work, please cite the following paper.

@article{wang2023learning,
  title={Learning to Hallucinate Face in the Dark},
  author={Wang, Yuanzhi and Lu, Tao and Yao, Yuan and Zhang, Yanduo and Xiong, Zixiang},
  journal={IEEE Transactions on Multimedia},
  year={2023},
  publisher={IEEE}
}