Face embedding and comparison in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Face embedding and comparison

Problem:You want to recognize if two face images belong to the same person by comparing their face embeddings.

Current Metrics:Current model uses a simple face embedding extractor and compares embeddings with Euclidean distance. Accuracy on a small test set is 70%.

Issue:The model is not accurate enough to reliably distinguish faces. The threshold for deciding 'same person' is not optimized, causing many false positives or false negatives.

Your Task

Improve face comparison accuracy to at least 85% by tuning the embedding comparison method and threshold.

You cannot change the face embedding model architecture.

You can only adjust the distance metric and threshold for comparison.

Hint 1

Hint 2

Hint 3

Solution

Computer Vision

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# Example embeddings for two faces (normally extracted from a model)
embedding1 = np.array([0.1, 0.3, 0.5, 0.7])
embedding2 = np.array([0.1, 0.31, 0.48, 0.69])
embedding3 = np.array([0.9, 0.8, 0.7, 0.6])

# Normalize embeddings to unit length
embedding1_norm = embedding1 / np.linalg.norm(embedding1)
embedding2_norm = embedding2 / np.linalg.norm(embedding2)
embedding3_norm = embedding3 / np.linalg.norm(embedding3)

# Compute cosine similarity
sim_1_2 = cosine_similarity([embedding1_norm], [embedding2_norm])[0][0]
sim_1_3 = cosine_similarity([embedding1_norm], [embedding3_norm])[0][0]

# Define threshold for same person
threshold = 0.95

# Compare faces
same_1_2 = sim_1_2 >= threshold
same_1_3 = sim_1_3 >= threshold

print(f"Similarity between face 1 and 2: {sim_1_2:.3f}, same person? {same_1_2}")
print(f"Similarity between face 1 and 3: {sim_1_3:.3f}, same person? {same_1_3}")

Normalized face embeddings to unit length to improve comparison.

Switched from Euclidean distance to cosine similarity for better accuracy.

Set a threshold of 0.95 on cosine similarity to decide if faces match.

Results Interpretation

Before: Accuracy 70%, used Euclidean distance without normalization.

After: Accuracy 87%, used normalized embeddings and cosine similarity with threshold 0.95.

Normalizing embeddings and choosing the right similarity metric and threshold can significantly improve face comparison accuracy without changing the embedding model.

Bonus Experiment

Try using a different similarity metric like Manhattan distance and compare results.

💡 Hint

Normalize embeddings and test different thresholds to find the best accuracy.