This project presents a Face Detector designed for detecting faces in images and associating them with previously seen identities. The system is built upon a VGGFace2 model, fine-tuned using ArcFace and Triplet Losses to enhance recognition performance.
The implementation provides:
- A face bank for storing and managing known identities.
- An application capable of detecting faces from images or a live webcam feed.
The project leverages a deep learning-based algorithm implemented with the PyTorch framework, achieving 97% classification accuracy on a subset of the CelebA dataset.
The goal of this project was to create a system capable of accurately detecting and recognizing faces, The main functionalities include:
- Identifing and locating faces in images or video streams.
- Aligning faces for the neural network.
- Match detected faces against a database of known identities.
- Managing known faces for recognition.
- Performing detection and recognition on live webcam input.
- Training: The model was trained using the CelebA dataset.
- Testing & Evaluation: A subset of CelebA was used for testing and performance metrics.
- Presentation: A custom Players Dataset was created for demonstration purposes, containing a few images of Kylian Mbappé and Cristiano Ronaldo.
⚠ Disclaimer: I do not claim ownership of the images used in the Players Dataset. They are included solely for demonstration purposes.
- Python 3
- PyTorch - for building and training deep learning models
- OpenCV – for image processing and feature detection
- NumPy – for efficient numerical operations
- Matplotlib – for visualization
- einops – for tensor manipulation
- Git/GitHub – for version control and collaboration
The face recognition system achieved 97% classification accuracy on a subset of the CelebA dataset, demonstrating high reliability in identifying known faces.
To further evaluate performance, we analyzed False Rejection Rate (FRR) and False Acceptance Rate (FAR) across different decision thresholds:
- FRR (False Rejection Rate) – Measures the percentage of genuine faces incorrectly rejected by the system.
- FAR (False Acceptance Rate) – Measures the percentage of impostor faces incorrectly accepted as known identities.
The following figure presents the FRR and FAR curves, showing the trade-off between security and recognition accuracy:
A well-balanced threshold ensures both low rejection of genuine users and high resistance to false acceptances.
- Jakub Bednarski – Conceptualization, Methodology, Software Development, Project Administration
- Julia Komorowska – Software Development, Investigation
- Adam Wasiela – Software Development, Investigation
- Hubert Woziński – Software Development, Investigation