Face Embeddings

Face embeddings are mathematical representations of faces that enable comparison and matching. This page explains how embeddings work in Visage.

What is a Face Embedding?

A face embedding is a fixed-length numerical vector that represents the unique features of a face. The key property is that similar faces produce similar embeddings.

Properties

• Fixed Length: Always the same dimension (128, 512, or 2622 depending on model)
• Normalized: L2-normalized to unit length
• Invariant: Relatively consistent across lighting and minor pose changes
• Comparable: Can be compared using cosine similarity or Euclidean distance

Available Models

Facenet (Default)

• Dimensions: 128
• Speed: Fast
• Accuracy: Good
• Use Case: General purpose, default choice

VGG-Face

• Dimensions: 2622
• Speed: Slower
• Accuracy: Very good
• Use Case: When accuracy is critical

ArcFace

• Dimensions: 512
• Speed: Medium
• Accuracy: State-of-the-art
• Use Case: Best accuracy/speed tradeoff

OpenFace

• Dimensions: 128
• Speed: Very fast
• Accuracy: Moderate
• Use Case: Resource-constrained environments

How Embeddings Work

1. Face Preprocessing

Detected face region is:

1. Aligned based on eye positions
2. Resized to model input size
3. Normalized pixel values

2. Neural Network Forward Pass

The preprocessed face is passed through a deep neural network trained to:

• Map faces to embedding space
• Ensure similar faces cluster together
• Maintain distance between different people

3. L2 Normalization

The raw output is normalized to unit length:

normalized_embedding = embedding / np.linalg.norm(embedding)

Similarity Calculation

Visage uses cosine similarity to compare embeddings:

similarity = np.dot(embedding1, embedding2)

Since embeddings are L2-normalized, this simplifies to the dot product.

Range: 0.0 (completely different) to 1.0 (identical)

Configuration

Set the embedding model in your .env file:

FACE_MODEL=Facenet  # or VGG-Face, ArcFace, OpenFace

Model Comparison

Model	Dimensions	Speed	Accuracy	Model Size
Facenet	128	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	~100MB
VGG-Face	2622	⭐⭐	⭐⭐⭐⭐⭐	~550MB
ArcFace	512	⭐⭐⭐⭐	⭐⭐⭐⭐⭐	~250MB
OpenFace	128	⭐⭐⭐⭐⭐	⭐⭐⭐	~30MB

Best Practices

1. Consistent Model: Use the same model for registration and identification
2. Multiple Samples: Register 3-5 embeddings per person
3. Good Quality: Use well-lit, frontal face images
4. Regular Updates: Re-register if appearance changes significantly

Next Steps

• System Overview - Return to system overview
• API Reference - Explore the API