Face embedding and comparison in Computer Vision - Model Pipeline Trace

Model Pipeline - Face embedding and comparison

This pipeline takes face images, turns them into special number lists called embeddings, and then compares these lists to see how similar the faces are. It helps in recognizing or verifying people by their faces.

Data Flow - 4 Stages

1Input Image

1 image x 160 x 160 x 3 (RGB)→Receive a face image of size 160x160 pixels with 3 color channels→1 image x 160 x 160 x 3

A photo of a person's face with height and width 160 pixels

↓

2Face Detection and Alignment

1 image x 160 x 160 x 3→Detect face location and align it to a standard pose→1 aligned face image x 160 x 160 x 3

Face centered and rotated so eyes are horizontal

↓

3Embedding Extraction

1 aligned face image x 160 x 160 x 3→Pass image through a neural network to get a 128-dimensional vector→1 vector x 128 dimensions

[0.12, -0.05, 0.33, ..., 0.07] (128 numbers representing the face)

↓

4Embedding Comparison

2 vectors x 128 dimensions→Calculate distance (e.g., Euclidean) between two face embeddings→1 similarity score (float)

Distance = 0.45 (lower means more similar)

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.5 |*
0.35| 
Epochs ->

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning to map faces to embeddings
2	0.9	0.60	Loss decreases, accuracy improves as embeddings get better
3	0.7	0.72	Model learns more distinct embeddings for different faces
4	0.5	0.82	Clear improvement in face representation quality
5	0.35	0.90	Model converges with good accuracy on face similarity

Prediction Trace - 4 Layers

Layer 1: Input Image

Layer 2: Face Detection and Alignment

Layer 3: Neural Network Embedding Layer

Layer 4: Embedding Comparison

Model Quiz - 3 Questions

What does the embedding vector represent in this pipeline?

A The original face image pixels

B The distance between two faces

C A list of numbers capturing unique face features

D The aligned face image

Key Insight

Face embedding converts complex face images into simple number lists that computers can compare easily. Aligning faces first helps the model focus on important features. Training improves the model so embeddings for the same person get closer, and for different people get farther apart.