PyTorch Face Detection from Scratch

This project is a challenge to myself to be able to learn and implement face detection from scratch. To do this I use the following Python libraries:

• PyTorch
• OpenCV
• PyTorch Lightning
• Albumentations

I used original Yolo Paper as a reference to implement the algorithms involved.

I experimented with different backbones and loss function hyperparameters (such as penalty weighting for incorrect predictions). From my testing I concluded several points:

• The Yolo loss function performs poorly on very small objects close together (e.g. a crowd of faces).
• The Yolo loss function performs well on larger objects (e.g. a single face).
• The Yolo loss function is very sensitive to over or under tuning the hyperparameters.

The more interesting part I found about the project was trying to get the model as small as possible, while maintaining the same accuracy. For this I tried 3 different architectures:

• A standard Resnet backbone, to get a good starting point.
• A Pool Resnet backbone, to reduce the computation time, without changing the parameters.
• A pretrained Mobilenet v3 backbone, to test if pretraining made any different.

I found that:

• The standard Resnet backbone performs well, but can be too slow when the number of bounding boxes is greater than 100.
• The Pool Resnet backbone performs equally well, but is much faster.
• The Mobilenet v3 backbone performs the same as the Pool Resnet, leading me to believe that none of the pretraining is helping.

Run the model

To run the medium PoolResnet Model:

1. Install the requirements

pip install -r requirements.txt

2. Run the model

python demo_model.py

Train a model

(Note: you'll need to install pytorch with cuda for training) To train the medium PoolResnet Model:

1. Install the requirements

pip install -r requirements.txt

2. Run the model

python train_model.py