Example of how to build a Face Detector with TF Lite C++ API on Android

This repo demonstrates how to use TF Lite to build a face detector with Java Native Interface (JNI). The pretrained CenterFace model was used. Model size has been reduced from 7.18MB to 2.3MB (by 68%).

Repo structure

app/src/main/

├── assets
│   └── centerface_w640_h480.tflite
├── cpp
│   ├── CMakeLists.txt
│   ├── face-detection.cpp
│   ├── face-detection.h
│   ├── main-native-lib.cpp
│   └── tf-lite-api
│       ├── README.md
│       ├── generated-libs
│       │   ├── arm64-v8a
│       │   │   └── libtensorflowlite.so
│       │   └── armeabi-v7a
│       │       └── libtensorflowlite.so
│       ├── include
│       │   ├── abseil
│       │   ├── flatbuffers
│       │   └── tensorflow
│       │       └── lite
│       └── run_me.ipynb
└── java
    └── cuongvng
        └── facedetection
            └── MainActivity.java

• app/src/main/assets contains the TF Lite model centerface_w640_h480.tflite. The original ONNX model was converted to TF Lite format (converting flow: ONNX -> TF graph -> TF Lite).
• app/src/main/cpp: core functions of the app
  • face-detection.h and face-detection.cpp are the header and source files which implement the detecting functions
  • main-native-lib.cpp is the JNI file that wrap the core functions to call them from Java.
  • tf-lite-api/ contain the shared libraries (.so) and headers of TF Lite C++, including dependencies such as abseil and flatbuffers. Details of how to build the libraries can be found on this repo.
• app/src/main/java/cuongvng/facedetection/MainActivity.java: High-level Android file, capturing frames from camera, call core functions via JNI.

Results: