Project Video
Watch the full project presentation and demo on YouTube.
Interactive Demo
Click faces to compare embeddings. Watch similarity scores and clustering in the embedding space.
About This Project
This project implements a deep face recognition pipeline using convolutional neural networks to learn discriminative face embeddings. Faces are mapped into a compact 128-dimensional embedding space where Euclidean distance directly corresponds to facial similarity. The system leverages triplet loss training to ensure that embeddings of the same identity cluster tightly while different identities remain well-separated.
The interactive demo above visualizes the core concept: procedurally generated abstract faces are each assigned embedding vectors, and clicking on faces reveals the cosine similarity between them. The 2D embedding space view uses t-SNE-style dimensionality reduction to show how similar faces naturally cluster together, demonstrating the power of learned metric spaces for recognition tasks.
Embedding Network
Deep CNN architecture maps facial images into a 128-dimensional compact representation space optimized via triplet loss.
Metric Learning
Cosine similarity and Euclidean distance metrics enable identity verification with configurable thresholds.
Face Clustering
Unsupervised clustering in embedding space groups unlabeled faces by identity using DBSCAN and hierarchical methods.
Real-Time Inference
Optimized model achieves sub-100ms face encoding, enabling real-time verification and identification workflows.