InspireFace

InspireFace is a cross-platform face recognition SDK developed in C/C++, supporting multiple operating systems and various backend types for inference, such as CPU, GPU, and NPU.

If you require further information on tracking development branches, CI/CD processes, or downloading pre-compiled libraries, please visit our development repository.

Please contact contact@insightface.ai using your company e-mail for commercial support, including obtaining and integrating higher accuracy models, as well as custom development.

---

📘 Documentation is a work in progress.
We welcome your questions💬, they help guide and accelerate its development.

Change Logs

2025-06-15 The ErrorCode-Table has been reorganized and streamlined.

2025-06-08 Add facial expression recognition.

2025-04-27 Optimize some issues and provide a stable version.

2025-03-16 Acceleration using NVIDIA-GPU (CUDA) devices is already supported.

2025-03-09 Release of android sdk in JitPack.

2025-02-20 Upgrade the face landmark model.

2025-01-21 Update all models to t3 and add tool to convert cosine similarity to percentage.

2025-01-08 Support inference on Rockchip devices RK3566/RK3568 NPU.

2024-12-25 Add support for optional RKRGA image acceleration processing on Rockchip devices.

2024-12-22 Started adapting for multiple Rockchip devices with NPU support, beginning with RV1103/RV1106 support.

2024-12-10 Added support for quick installation via Python package manager.

2024-10-09 Added system resource monitoring and session statistics.

2024-09-30 Fixed some bugs in the feature hub.

2024-08-18 Updating Benchmark: Using CoreML with Apple's Neural Engine (ANE) on the iPhone 13, the combined processes of Face Detection + Alignment + Feature Extraction take less than 2ms.

2024-07-17 Add global resource statistics monitoring to prevent memory leaks.

2024-07-07 Add some face action detection to the face interaction module.

2024-07-05 Fixed some bugs in the python ctypes interface.

2024-07-03 Add the blink detection algorithm of face interaction module.

2024-07-02 Fixed several bugs in the face detector with multi-level input.

License

The licensing of the open-source models employed by InspireFace adheres to the same requirements as InsightFace, specifying their use solely for academic purposes and explicitly prohibiting commercial applications.

Quick Start

For Python users on Linux and MacOS, InspireFace can be quickly installed via pip:

pip install inspireface

After installation, you can use inspireface like this:

import cv2
import inspireface as isf

# Create a session with optional features
opt = isf.HF_ENABLE_NONE
session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)

# Load the image using OpenCV.
image = cv2.imread(image_path)

# Perform face detection on the image.
faces = session.face_detection(image)

for face in faces:
    x1, y1, x2, y2 = face.location
    rect = ((x1, y1), (x2, y2), face.roll)
     # Calculate center, size, and angle
    center = ((x1 + x2) / 2, (y1 + y2) / 2)
    size = (x2 - x1, y2 - y1)
    angle = face.roll

    # Apply rotation to the bounding box corners
    rect = ((center[0], center[1]), (size[0], size[1]), angle)
    box = cv2.boxPoints(rect)
    box = box.astype(int)

    # Draw the rotated bounding box
    cv2.drawContours(image, [box], 0, (100, 180, 29), 2)

cv2.imshow("face detection", image)
cv2.waitKey(0)
cv2.destroyAllWindows()

⚠️The project is currently in a rapid iteration phase, before each update, please pull the latest model from the remote side!

import inspireface

for model in ["Pikachu", "Megatron"]:
    inspireface.pull_latest_model(model)

More examples can be found in the python directory.

Preparation

Clone 3rdparty

Clone the 3rdparty repository from the remote repository into the root directory of the project. Note that this repository contains some submodules. When cloning, you should use the --recurse-submodules parameter, or after entering the directory, use git submodule update --init --recursive to fetch and synchronize the latest submodules:

# Must enter this directory
cd InspireFace
# Clone the repository and pull submodules
git clone --recurse-submodules https://github.com/tunmx/inspireface-3rdparty.git 3rdparty

If you need to update the 3rdparty repository to ensure it is current, or if you didn't use the --recursive parameter during the initial pull, you can run git submodule update --init --recursive:

# Must enter this directory
cd InspireFace
# If you're not using recursive pull
git clone https://github.com/tunmx/inspireface-3rdparty.git 3rdparty

cd 3rdparty
git pull
# Update submodules
git submodule update --init --recursive

Downloading Model Package Files

You can download the model package files containing models and configurations needed for compilation from Release Page and extract them to any location.

You can use the command/download_models_general.sh command to download resource files, which will be downloaded to the test_res/pack directory. This way, when running the Test program, it can access and read the resource files from this path by default.

⚠️The project is currently in a rapid iteration phase, before each update, please pull the latest model from the remote side!

# Download lightweight resource files for mobile device
bash command/download_models_general.sh Pikachu
# Download resource files for mobile device or PC/server
bash command/download_models_general.sh Megatron
# Download resource files for RV1109
bash command/download_models_general.sh Gundam_RV1109
# Download resource files for RV1106
bash command/download_models_general.sh Gundam_RV1106
# Download resource files for RK356X
bash command/download_models_general.sh Gundam_RK356X
# Download resource files for RK3588
bash command/download_models_general.sh Gundam_RK3588
# Download resource files for NVIDIA-GPU Device(TensorRT)
bash command/download_models_general.sh Megatron_TRT

# Download all model files
bash command/download_models_general.sh

Installing MNN

The '3rdparty' directory already includes the MNN library and specifies a particular version as the stable version. If you need to enable or disable additional configuration options during compilation, you can refer to the CMake Options provided by MNN. If you need to use your own precompiled version, feel free to replace it.

Requirements

• CMake (version 3.20 or higher)
• NDK (version 16 or higher, only required for Android) [Optional]
• MNN (version 1.4.0 or higher)
• C++ Compiler
  • Either GCC or Clang can be used (macOS does not require additional installation as Xcode is included)
    • Recommended GCC version is 4.9 or higher
      • Note that in some distributions, GCC (GNU C Compiler) and G++ (GNU C++ Compiler) are installed separately.
      • For instance, on Ubuntu, you need to install both gcc and g++
    • Recommended Clang version is 3.9 or higher
  • arm-linux-gnueabihf (for RV1109/RV1126) [Optional]
    • Prepare the cross-compilation toolchain in advance, such as gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf
• CUDA (version 11.x or higher) [Optional]
  • GPU-based inference requires installing NVIDIA's CUDA dependencies on the device.
• TensorRT (version 10 or higher) [Optional]

• Eigen3

• RKNN [Optional]

  • Adjust and select versions currently supported for specific requirements.

Compilation

CMake option are used to control the various details of the compilation phase. Please select according to your actual requirements. CMake Option.

Local Compilation

If you are using macOS or Linux, you can quickly compile using the shell scripts provided in the command folder at the project root:

cd InspireFace/
# Execute the local compilation script
bash command/build.sh

After compilation, you can find the local file in the build directory, which contains the compilation results. The install directory structure is as follows:

inspireface-linux
   ├── include
   │   ├── herror.h
   │   ├── intypedef.h
   │   ├── inspireface.h
   │   ├── inspirecv/
   │   └── inspireface/
   └── lib
       └── libInspireFace.so

• libInspireFace.so：Compiled dynamic linking library.
• inspireface.h：Header file definition.
• herror.h：Reference error number definition.
• intypedef.h: Type definition file.
• inspirecv: Simple cv library CPP header file folder.
• inspireface: inspireface cpp header folder.

Cross Compilation

Cross compilation requires you to prepare the target platform's cross-compilation toolchain on the host machine in advance. Here, compiling for Rockchip's embedded devices RV1106 is used as an example:

# Set the path for the cross-compilation toolchain
export ARM_CROSS_COMPILE_TOOLCHAIN=YOUR_DIR/arm-rockchip830-linux-uclibcgnueabihf
# Execute the cross-compilation script for RV1106
bash command/build_cross_rv1106_armhf_uclibc.sh

After the compilation is complete, you can find the compiled results in the build/inspireface-linux-armv7-rv1106-armhf-uclibc directory.

iOS Compilation

To compile for iOS, ensure you are using a Mac device. The script will automatically download third-party dependencies into the .macos_cache directory.

bash command/build_ios.sh

After the compilation is complete, inspireface.framework will be placed in the build/inspireface-ios directory.

Android Compilation

You can compile for Android using the following command, but first you need to set your Android NDK path:

export ANDROID_NDK=YOUR_ANDROID_NDK_PATH
bash command/build_android.sh

After the compilation is complete, arm64-v8a and armeabi-v7a libraries will be placed in the build/inspireface-android directory.

Linux-based NVIDIA GPU Acceleration with TensorRT Compilation

If you want to use NVIDIA GPU devices for accelerated inference on Linux, you need to install CUDA, cuDNN, and TensorRT-10 on your device, and configure the relevant environment variables.

# Example, Change to your TensorRT-10 path
export TENSORRT_ROOT=/user/tunm/software/TensorRT-10

Before compiling, please ensure that your related environments such as CUDA and TensorRT-10 are available. If you encounter issues with finding CUDA libraries during the compilation process, you may need to check whether the relevant environment variables have been configured: CUDA_TOOLKIT_ROOT_DIR, CUDA_CUDART_LIBRARY.

bash command/build_linux_tensorrt.sh

Additionally, you can use NVIDIA's Docker images for compilation. For example, to compile using a CUDA 12 and TensorRT-10 image on Ubuntu 22.04, you can execute the following commands:

docker-compose up build-tensorrt-cuda12-ubuntu22

If you want to use pre-compiled libraries, you can use FindTensorRT.cmake to create links to CUDA and TensorRT.

React Native

For Android and iOS, in addition to the native interface, you can use the React Native library powered by Nitro Modules and JSI—providing ultra-fast, seamless bindings to the InspireFace SDK. For more details, check out the react-native-nitro-inspire-face repository or the documentation. Author: ronickg.

Supported Platforms and Architectures

We have completed the adaptation and testing of the software across various operating systems and CPU architectures. This includes compatibility verification for platforms such as Linux, macOS, iOS, and Android, as well as testing for specific hardware support to ensure stable operation in diverse environments.

| No. | Platform | Architecture

(CPU) | Device

(Special) | Supported | Passed Tests | Release

(Online) | | ------- | -------------------- | --------------------- | ------------