faceverify

Configuration Guide

Complete guide to configuring FaceVerify.

---

Configuration Methods

FaceVerify supports three configuration methods (in order of precedence):

1. Constructor arguments (highest priority)
2. Environment variables
3. YAML configuration file

---

Quick Configuration

Python

from faceverify import FaceVerifier
from faceverify.config import VerifierConfig

config = VerifierConfig(
    detector_backend="opencv",    # Fast detection
    embedding_model="deepface",   # Uses DeepFace with Facenet512
    threshold=0.65,               # Match threshold
)

verifier = FaceVerifier(config)

[!NOTE] VerificationConfig is an alias for VerifierConfig. Both work identically.

Environment Variables

export FACEVERIFY_DETECTOR=opencv
export FACEVERIFY_THRESHOLD=0.65

YAML File

# config.yaml
detector:
  backend: opencv
  confidence_threshold: 0.9

embedding:
  model: deepface

decision:
  threshold: 0.65

config = VerifierConfig.from_yaml("config.yaml")

---

Detector Backends

OpenCV (Default)

• Fastest option
• No extra installation
• Good for real-time applications

config = VerifierConfig(detector_backend="opencv")

MTCNN

• Better accuracy
• Requires: pip install mtcnn

config = VerifierConfig(detector_backend="mtcnn")

RetinaFace

• Best accuracy
• Slower than others
• Requires: pip install retinaface

config = VerifierConfig(detector_backend="retinaface")

MediaPipe

• Good balance of speed/accuracy
• Requires: pip install mediapipe

config = VerifierConfig(detector_backend="mediapipe")

[!TIP] Use opencv for speed, retinaface for accuracy.

---

Embedding Models

DeepFace (Default)

• Uses Facenet512 model internally
• 512-dimensional embeddings
• Best overall performance

config = VerifierConfig(embedding_model="deepface")

[!IMPORTANT] The embedding model uses DeepFace library with Facenet512. First run downloads model weights (~90MB).

Facenet

• Alias for deepface with Facenet512
• 512-dimensional embeddings

config = VerifierConfig(embedding_model="facenet")

---

Similarity Metrics

Cosine (Default)

• Range: 0 to 1 (higher = more similar)
• Best for normalized embeddings

config = VerifierConfig(similarity_metric="cosine")

Euclidean

• Range: 0 to infinity (lower = more similar)
• Sensitive to magnitude

config = VerifierConfig(similarity_metric="euclidean")

Manhattan

• Range: 0 to infinity (lower = more similar)
• Less sensitive to outliers

config = VerifierConfig(similarity_metric="manhattan")

---

Threshold Tuning

Understanding Thresholds

Threshold	Effect
Higher (0.7+)	Stricter matching, fewer false positives
Lower (0.5-)	Looser matching, fewer false negatives

Use Cases

Use Case	Recommended Threshold
High security	0.75 - 0.80
Standard verification	0.60 - 0.70
Loose matching	0.50 - 0.55

[!IMPORTANT] Always test with your specific dataset to find optimal threshold.

Setting Threshold

# At initialization
config = VerifierConfig(threshold=0.70)

# Or modify later
verifier.threshold = 0.70

---

Performance Settings

GPU Acceleration

config = VerifierConfig(enable_gpu=True)

[!NOTE] GPU requires CUDA-compatible TensorFlow installation.

Batch Size

config = VerifierConfig(batch_size=32)

Larger batch sizes improve throughput but use more memory.

---

Complete Example

```python from faceverify import FaceVerifier from faceverify.config import VerifierConfig

High-accuracy configuration

config = VerifierConfig( # Detection detector_backend=”retinaface”, detector_confidence=0.95,

# Embedding
embedding_model="deepface",
normalize_embeddings=True,

# Similarity
similarity_metric="cosine",

# Decision
threshold=0.70,

# Performance
enable_gpu=True,
batch_size=16, )

verifier = FaceVerifier(config)

This site is open source. Improve this page.