Face Recognition Models

Overview

degirum-face uses two types of AI models for face recognition: face detection models and face embedding models. Both are optimized and compiled for a wide range of hardware platforms through the DeGirum model registry.

Important: The degirum_face library provides code and pipelines for face recognition workflows. Model licensing is separate from the degirum_face library licensing. See the license section under each model type for specific licensing requirements.

Face Detection Models

Purpose

Face detection models locate faces in images and identify 5 facial keypoints (landmarks):

• Left eye

• Right eye

• Nose

• Left mouth corner

• Right mouth corner

These keypoints enable face alignment - a critical step for accurate embedding extraction.

License

Face detection models are trained by DeGirum and can be used commercially when users license the degirum-face package. Contact DeGirum for licensing information.

Model Details

• Training: All face detection models are trained by DeGirum

• Output: Bounding boxes + 5 keypoints per detected face

• Compilation: Models are compiled and optimized for all supported hardware platforms:

  • Hailo (HAILORT/HAILO8, HAILORT/HAILO8L)

  • Axelera (AXELERA/METIS)

  • DEEPX (DEEPX/M1A)

  • Intel (OPENVINO/CPU, OPENVINO/GPU, OPENVINO/NPU)

  • NVIDIA (TENSORRT/GPU)

  • Rockchip (RKNN/RK3588)

  • Google (TFLITE/EDGETPU)

  • DeGirum (N2X/ORCA1)

Usage

Detection models are automatically selected from the model registry based on your hardware configuration:

detector_spec = degirum_face.get_face_detection_model_spec(
    device_type="HAILORT/HAILO8",
    inference_host_address="@cloud"
)

Face Embedding Models

Purpose

Face embedding models convert aligned face images into numerical vectors (embeddings) that capture unique facial features. These embeddings enable face matching and identification.

License

The ONNX file for the face embedding model is from the Hailo Model Zoo and is licensed under the MIT license. See THIRD-PARTY-LICENSES.md for the complete license text and copyright notice.

Model Details

• Architecture: ArcFace MobileFacenet

• Source: ONNX model from Hailo Model Zoo

• Compilation: DeGirum compiled the ONNX model for all supported hardware platforms

• Output: 512-dimensional embedding vector per face

ArcFace Model

The ArcFace (Additive Angular Margin Loss) model is a state-of-the-art face recognition architecture that:

• Uses MobileFacenet backbone for feature extraction

• Produces highly discriminative face embeddings

• Achieves excellent accuracy on face recognition benchmarks

• Generalizes well across different lighting conditions, poses, and demographics

Usage

Embedding models are automatically selected from the model registry:

embedding_spec = degirum_face.get_face_embedding_model_spec(
    device_type="HAILORT/HAILO8",
    inference_host_address="@cloud"
)

Model Registry

The DeGirum model registry automatically selects the optimal model for your hardware platform. This eliminates the need to manually choose models or tune parameters.

Benefits:

• Pre-optimized models for each hardware platform

• Automatic model selection based on device type

• Consistent performance across different hardware

• Simplified configuration

Accessing the registry:

# List all supported hardware platforms
supported_hw = degirum_face.model_registry.get_hardware()

# Get available hardware on cloud or local
available_hw = degirum_face.get_system_hw("@cloud")

# Find compatible hardware you can use now
compatible_hw = degirum_face.get_compatible_hw("@cloud")

Custom Models

While the model registry provides optimized models for all supported platforms, you can also use custom models by creating ModelSpec objects directly. This is useful for:

• Using proprietary face detection models

• Testing alternative embedding architectures

• Integrating domain-specific models

See ModelSpec Documentation for details on using custom models.

Model Performance

Performance varies by hardware platform and model complexity. Key factors:

Face Detection:

• Input resolution impacts speed and accuracy

• Higher resolution = better small face detection, slower processing

• Typical throughput: 10-100+ FPS depending on hardware

Face Embedding:

• Fixed input size (typically 112×112 pixels)

• Extraction time: 1-50ms per face depending on hardware

• Batch processing improves throughput for multiple faces

For production deployments, test with your specific hardware and video resolution to determine optimal settings.

Acknowledgments

We gratefully acknowledge Hailo for providing the ArcFace MobileFacenet ONNX model through the Hailo Model Zoo. This model enables high-quality face recognition capabilities in degirum-face. DeGirum has compiled and optimized this model for all supported hardware platforms to ensure consistent performance across different deployment scenarios.