BlazeFace

An implementation of google's BlazeFace paper using Python 3.9.6.

For an autonomous drone competition my team needed a computer vision model capable of running fast on SBC's like Raspberry Pi and detecting geometric figures on the ground. To this end I've created a couple of modules:

• polygons.py: A simple module for generating and inpainting geometric figures on images.
• data.py: Integration with PyTorch datasets
• model.py: The BlazeFace model itself and it's building blocks
• losses.py: Implementation of YOLO losses
• train.py: The training code
• quick_eval.py: Visualizing the results
• utils.py

Instructions

Install requirements.txt, change the paths in the code and run using python -m blaze.train or interactively using vscode.