Light Speed Face Detection (YuNet, MMDetection)

Yunet is one of the fastest (maybe lightest) face detection in the world with only 75M params but achieve pretty well performance. There are some thing we can improve on top of it and this repo is to do so.

This repo strongly based on Yunet training repo.

Updated:

Since 3 years from the lastest update for Yunet tranining repo, in order to install necessary libs please use the following sets:

Fresh venv recommended

1. Install torch <= I'm using Cuda 12.9 - NVIDIA 3090

 pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 \
 -f https://download.pytorch.org/whl/torch_stable.html

2. Install mmcv

 pip install -U openmim
 mim install "mmcv-full==1.6.0" \
 -f https://download.openmmlab.com/mmcv/dist/cu116/torch1.12/index.html

Improvement

1. Use takenique SE - squeeze and excitation block
2. Increase the size of model from 'nano' to 'tiny', it's able to increase to medium or large if we need but trade-off for x% speed. But the speed is already high.
   Note: WiderFace is considered as a big dataset for face detection, it's common to increase the size (tradeoff for speed) to gain higher metrics
3. Increase batch size training from 16 to 32 or higher also contribute to the mAP
4. I also add other dataset like: face in COCO2017, Face Detection Kaggle, Lagenda but it seems no improvement due to the different in data distribution (Maybe)

Training

CUDA_VISIBLE_DEVICES=0,1 bash tools/dist_train.sh ./configs/yunet_n.py 2 12345

Testing on Wider face

python tools/test_widerface.py ./configs/yunet_n.py ./weights/yunet_n.pth --mode 2

Result

Baseline Yunet_n: AP_easy=0.892, AP_medium=0.883, AP_hard=0.811
SE Yunet_n: AP_easy=0.892, AP_medium=0.884, AP_hard=0.815    | only + about 10% params
Tiny Yunet_t: AP_easy=0.906, AP_medium=0.896, AP_hard=0.829  | + about 30% params
Medium Yunet_m: AP_easy=0.913, AP_medium=0.90338, AP_hard=0.838 | + about 50% params