FastFace: Lightweight Face Detection Framework

Easy-to-use face detection framework, developed using pytorch-lightning.
Checkout documentation for more.

Key Features

• 🔥 Use pretrained models for inference with just few lines of code
• 📈 Evaluate models on different datasets
• 🛠️ Train and prototype new models, using pre-defined architectures
• 🚀 Export trained models with ease, to use in production

Contents

• Installation
• Pretrained Models
• Demo
• Benchmarks
• Tutorials
• References
• Citations

Installation

From PyPI

pip install fastface -U

From source

git clone https://github.com/borhanMorphy/fastface.git
cd fastface
pip install .

Pretrained Models

Pretrained models can be accessable via fastface.FaceDetector.from_pretrained(<name>)

Name	Architecture	Configuration	Parameters	Model Size	Link
lffd_original	lffd	original	2.3M	9mb	weights
lffd_slim	lffd	slim	1.5M	6mb	weights

Demo

Using package

import fastface as ff
import imageio
from pytorch_lightning.utilities.model_summary import ModelSummary

# load image as RGB
img = imageio.imread("<your_image_file_path>")[:,:,:3]

# build model with pretrained weights
model = ff.FaceDetector.from_pretrained("lffd_original")
# model: pl.LightningModule

# get model summary
ModelSummary(model, max_depth=1)

# set model to eval mode
model.eval()

# [optional] move model to gpu
model.to("cuda")

# model inference
preds, = model.predict(img, det_threshold=.8, iou_threshold=.4)
# preds: {
#    'boxes': [[xmin, ymin, xmax, ymax], ...],
#    'scores':[<float>, ...]
# }

Using demo.py script

python demo.py --model lffd_original --device cuda --input <your_image_file_path>

sample output;

Benchmarks

Following results are obtained with this repository

WIDER FACE

validation set results

Name	Easy	Medium	Hard
lffd_original	0.893	0.866	0.758
lffd_slim	0.866	0.854	0.742

Tutorials

• Widerface Benchmark
• BentoML Deployment

References

• LFFD Paper
• Official LFFD Implementation

Citations

@inproceedings{LFFD,
    title={LFFD: A Light and Fast Face Detector for Edge Devices},
    author={He, Yonghao and Xu, Dezhong and Wu, Lifang and Jian, Meng and Xiang, Shiming and Pan, Chunhong},
    booktitle={arXiv:1904.10633},
    year={2019}
}