InspireFace: Cross-Platform C/C++ Face Recognition SDK for Edge Devices

Introduction

We are excited to announce the release of InspireFace, a high-performance C/C++ face recognition SDK designed for cross-platform edge deployment. InspireFace brings InsightFace's state-of-the-art face analysis capabilities to mobile, desktop, and embedded devices with a unified, easy-to-use API.

Key Features

Extreme Performance

InspireFace is engineered for speed:

• Full pipeline (detection + alignment + feature extraction) in under 2ms on iPhone 13 (CoreML/ANE)
• Optimized for CPU, GPU, and NPU inference
• Lightweight models suitable for resource-constrained devices

Cross-Platform Support

Deploy on virtually any platform:

• Linux: x86_64, ARM64, RISC-V
• macOS: Intel and Apple Silicon
• iOS: iPhone and iPad with CoreML acceleration
• Android: ARM and x86 with NNAPI support
• Embedded: Rockchip NPU (RK3588/RK3566/RK3568)
• CUDA: NVIDIA GPU acceleration for server deployment

Comprehensive Functionality

InspireFace provides a complete face analysis pipeline:

• Face Detection: High-accuracy face localization
• Face Recognition: ArcFace-based identity embedding extraction
• Face Alignment: Precise landmark detection
• Attribute Analysis: Age, gender, and expression estimation
• Liveness Detection: Anti-spoofing with presentation attack detection
• Head Pose Estimation: Yaw, pitch, and roll estimation
• Facial Action Detection: Blink, nod, and shake detection
• Mask Detection: Face mask presence detection

Architecture

InspireFace is built with a modular, layered architecture:

1. C API Layer: Stable ABI for language interoperability

2. Core Engine: Optimized inference pipeline

3. Backend Abstraction: Pluggable inference backends (ONNX Runtime, CoreML, TensorRT, RKNN)

4. Model Management: Efficient model loading and caching

Quick Start

C/C++ Integration

#include "inspireface.h"

// Initialize
HResult ret = HFLaunchInspireFace(model_path);

// Create session
HFSession session;
HFCreateInspireFaceSession(
    HF_ENABLE_FACE_RECOGNITION | HF_ENABLE_LIVENESS,
    HF_DETECT_MODE_ALWAYS_DETECT,
    3,  // max face count
    -1, // use default pixel format
    0,  // use default thread count
    &session
);

// Process image
HFImageData image = {width, height, 3, data};
HFImageStream stream;
HFCreateImageStream(&image, &stream);

// Detect and recognize
HFMultipleFaceData faces;
HFExecuteFaceTrack(session, stream, &faces);
HFFaceFeature feature;
HFExtractFaceFeature(session, stream, faces.tokens[0], &feature);

Python Integration

import cv2
import inspireface as isf

# Initialize and create session
session = isf.InspireFaceSession(
    isf.HF_ENABLE_FACE_RECOGNITION | isf.HF_ENABLE_LIVENESS,
    isf.HF_DETECT_MODE_ALWAYS_DETECT
)

# Process image
image = cv2.imread("face.jpg")
faces = session.face_detection(image)

# Extract features
for face in faces:
    feature = session.face_feature_extract(image, face)
    print(f"Feature dimension: {len(feature)}")

Performance Benchmarks

Licensing

InspireFace is available under two licensing options:

• Academic License: Free for research and educational use
• Commercial License: Contact us for production deployment

Getting Started

Visit the InspireFace GitHub repository for documentation, examples, and pre-built binaries. For commercial licensing inquiries, contact us at contact@insightface.ai.