insightface-rk 0.7.3rc1

Enhanced face analysis package with frontal face detection and filtering largest face / small faces

InsightFace-Rk

An enhanced face analysis package built on top of InsightFace, with additional features for frontal face detection, largest face filtering, and improved gender/age estimation.

Features

• All features from InsightFace base package
• Enhanced frontal face detection with angle validation
• Automatic largest face filtering
• Improved gender and age estimation
• Configurable face size thresholds

Installation

pip install insightface-rk

Usage

from insightface.app import FaceAnalysis

# Initialize face analyzer with enhanced features
face_analyzer = FaceAnalysis(
    name='buffalo_l',
    root='.insightface',
    providers=['CUDAExecutionProvider', 'CPUExecutionProvider'],
    allowed_modules=['detection', 'recognition', 'genderage']
)

# Prepare the model with custom thresholds
face_analyzer.prepare(
    ctx_id=0,
    det_size=(960, 960),
    det_thresh=0.7
)

# Get faces with enhanced filtering
faces = face_analyzer.get(
    img,
    max_num=1,                    # Maximum number of faces to detect
    small_face_threshold=0.05,    # Minimum relative face size
    largest_face_only=True,       # Only return the largest face
    genderage=True               # Enable gender and age estimation
)

# Process detected faces
for face in faces:
    print(f"Age: {face.age}")
    print(f"Gender: {'Male' if face.gender == 1 else 'Female'}")
    print(f"Detection score: {face.det_score}")

Enhanced Features Explained

Frontal Face Detection

The package includes additional validation to ensure detected faces are sufficiently frontal-facing, improving the accuracy of subsequent recognition tasks.

Largest Face Filtering

When largest_face_only=True, the system will automatically select the largest detected face in the image, useful for processing portrait photos or when the main subject should be prioritized.

Face Size Thresholding

The small_face_threshold parameter allows filtering out faces that are too small relative to the image dimensions, helping to focus on the most relevant faces in the scene.

Credits

This package is built on top of InsightFace (v0.7.3), an excellent face analysis toolkit. We extend our gratitude to the InsightFace team for their foundational work.

License

This package is released under the MIT License. However, please note that the pretrained models from InsightFace are available for non-commercial research purposes only.

Requirements

• Python >= 3.8
• CUDA-compatible GPU (optional, but recommended for better performance)
• See requirements.txt for full dependencies