This is unofficial tensorflow blazeface implementation from scratch. This repo includes the entire training pipeline of blazeface. However, since the dataset used in the training process is a modified version of some datasets, it is not shared at this stage. Anchor / prior box hyperparameters were taken from the MediaPipe implementation. Loss calculation and augmentation methods were implemented as in SSD.

It's implemented and tested with tensorflow 2.0, 2.1, and 2.2

Project models created in virtual environment using miniconda. You can also create required virtual environment with conda.

To create virtual environment (tensorflow-2 gpu environment):

conda env create -f environment.yml

To train and test BlazeFace model:

python trainer.py
python predictor.py

If you have GPU issues you can use -handle-gpu flag with these commands:

python trainer.py -handle-gpu

• BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs [paper]
• SSD: Single Shot MultiBox Detector [paper]
• MediaPipe [code]
• BlazeFace-PyTorch [code]