The official PyTorch implementation for HLA-Face: Joint High-Low Adaptation for Low Light Face Detection (CVPR21).

You can find more information on our project website.

Our face detection mAP on DARK FACE test set (trained without DARK FACE labels): 44.4

Precision-Recall (PR) curves on DARK FACE.

Please follow [Training Code] to download and prepare the dataset.

Basically, to train the final Joint High-Low Adaptation framework

cd ./train_code
python train.py --multigpu

More details are given in [Training Code].

First, download and cd ./test_code.

• Python 3

• PyTorch 1.2.0

  • Higher versions may cause the RuntimeError: Legacy autograd function with non-static forward method is deprecated. Please use new-style autograd function with static forward method bug.

  • According to https://github.com/yxlijun/DSFD.pytorch, PyTorch 0.3 is also OK, but we have not checked it.

• opencv

• numpy

• easydict

Download the checkpoints from [Google] [Baidu (xnb6)], and save them as:

./final_weights/Illumination-Enhancer.pth
./final_weights/Face-Detector.pth

This script detects the example test.png image.

python test.py

The detection result (a text file) can be found in ./result/ and visualized by draw_result.py

Here, the left image is test.png, and the right image is the visualized face detection result.

You can further edit the function load_images() in test.py to load your testing images. The detection results can be found in ./result where each detection result is named by the name of the testing image.

For example, to test the DSFD test dataset:

def load_images():
    import glob
    return glob.glob('./YOUR_PATH_TO_DARKFACE/images/test/*.png')

If you want to run without the multi-scale testing scheme (the mode we used in ablation studies), set:

USE_MULTI_SCALE = False
MY_SHRINK = 2

This code is based on DSFD. Thanks a lot for the great work!