Face Clusters

In this repository, I've experimented with face recognition by clustering image embeddings. The image embeddings are generated using Retina-Face and reduced to a 2 or 3 dimensional representation using UMAP dimensionality reduction algorithm. Later the vectors are labelled into clusters using DBSCAN and HDBSCAN algorithms.

Project structure

.
├── clean.py
├── embed.py
├── faces
│   ├── face1
│   └── face2
├── faces.py
├── flow.sh
├── imgs
│   ├── face1
│   └── face2
├── other-methods
│   ├── faces-mtcnn.py
│   ├── faces-nudenet.py
│   └── faces-retina.py
├── README.md
├── requirements.txt
└── viz.py

Usage

• Clone the repo.
• Run pip install -r requirements.txt.
• Download the face dataset you like to use and set it according to the directory structure.
• Run the python files in this order (Or just run flow.sh):

python3 clean.py
python3 faces.py
python3 embed.py

Conclusions

Face Extraction

• MTCNN doesn't give much good results
• NudeNet is good but wastes compute on detecting other things
• retina-face is by far the best. Just uses a lot of resources. (Works good on Google Colab but, not recommended for local machine)

Face recognition

• Facenet works just right