blazeface

BlazeFace - Real time object detection in browser

This is the code for the Medium Article "BlazeFace: How to Run Real-time Object Detection in the Browser".

This has two separate folders:

• Model training with Python
• Model inference and demo with JavaScript

Model training with Python

All the commands in this section are expected to be used in the python folder of this repo.

Environment setup

You can then create a new environment and install all the required packages:

conda create -y -n blazeface python=3.10
conda activate blazeface
pip install -r python/requirements.txt

Dataset

If you're willing to train the model yourself, you need to download and build the dataset first.

To do so, you just need to run the following script:

python build_dataset.py

This will download the dataset, compute the expected label format, and separate it into train and validation sets. This will use the validation set of the Open Images Dataset, by Google.

More specifically, it will select only pictures with labeled human faces and a permissive enough license.

Model training

You can then train the BlazeFace model for 100 epochs:

python trainer.py --epochs 100

This will create a file weights/blazeface.pt with the trained weights of the model. Having a GPU is recommended, but not mandatory. Note that a pretrained TFLite model is available in javascript/assets/.

N.B.: check the help of the trainer.py script for more about the available parameters

If during training you encounter the following error RuntimeError: received 0 items of ancdata, just run ulimit -n 1048576 in your terminal, and then rerun the training script.

Convert model to TFLite

You can finally convert the model to TFLite:

python tf_lite_converter.py

This will create a weights/blazeface.tflite: the trained model converted to TFLite format.

As for the training script, you can check the help of the tf_lite_converter.py script for more about the available parameters

Face detection demo in the browser

This is allowing you to test the real time face detection in browser.

Copy/paste your converted model

If you want to reuse the provided model of this repo, just go to the next step.

If you want to use your trained model, make sure you are in the javascript folder of this repo, and then copy/paste the trained model:

cp ../python/weights/blazeface.tflite assets/.

Running the web demo

Just launch the index.html with your favorite web server (you can use you IDE such as PyCharm of VSCode).