GPU Accelerated TensorFlow Lite applications on Android NDK.

Run and measure the performance of TensorFlow Lite GPU Delegate on Android NDK.

1. Applications

Blazeface

• Lightweight Face Detection.
  

DBFace

• Higher accurate Face Detection.
  

Age Gender Estimation

• Detect faces and estimage their Age and Gender
• based on pretrained model of https://github.com/yu4u/age-gender-estimation
  

Image Classification

• Image Classfication using Moilenet.
  

Object Detection

• Object Detection using MobileNet SSD.
  

Hair Segmentation

• Hair segmentation and recoloring.
  

3D Handpose

• 3D Handpose Estimation from single RGB images.
  

Iris Detection

• Eye position estimation by detecting the iris.
  

Posenet

• Pose Estimation.
  

Depth Estimation (DenseDepth)

• Depth Estimation from single images.
  
• based on pretrained model of https://github.com/ialhashim/DenseDepth

Semantic Segmentation

• Assign semantic labels to every pixel in the input image.
  

Selfie to Anime

• Generate anime-style face image.
  
• based on pretrained model of https://github.com/margaretmz/Selfie2Anime-with-TFLite
  

Anime GAN

• Transform photos into anime style images.
  
• based on pretrained model of https://github.com/TachibanaYoshino/AnimeGANv2
  

U^2-Net portrait drawing

• Human portrait drawing by U^2-Net.
  

Artistic Style Transfer

• Create new artworks in artistic style.
  

MIRNet

• Enhance low-light images upto a great extent.
  
• based on pretrained model of https://github.com/sayakpaul/MIRNet-TFLite

2. How to Build & Run

2.1 setup environment

• Download and install Android NDK.

$ mkdir ~/Android/
$ mv ~/Download/android-ndk-r20b-linux-x86_64.zip ~/Android
$ cd ~/Android
$ unzip android-ndk-r20b-linux-x86_64.zip

• Download and install bazel.

$ wget https://github.com/bazelbuild/bazel/releases/download/3.1.0/bazel-3.1.0-installer-linux-x86_64.sh
$ chmod 755 bazel-3.1.0-installer-linux-x86_64.sh
$ sudo ./bazel-3.1.0-installer-linux-x86_64.sh

2.2 build TensorFlow Lite library and GPU Delegate library

• run the build script to build TensorFlow Library

$ mkdir ~/work
$ git clone https://github.com/terryky/android_tflite.git
$ cd android_tflite/third_party/
$ ./build_libtflite_r2.4_android.sh

(Tensorflow configure will start after a while. Please enter according to your environment)


$ ls -l tensorflow/bazel-bin/tensorflow/lite/
-r-xr-xr-x  1 terryky terryky 3118552 Dec 26 19:58 libtensorflowlite.so*

$ ls -l tensorflow/bazel-bin/tensorflow/lite/delegates/gpu/
-r-xr-xr-x 1 terryky terryky 80389344 Dec 26 19:59 libtensorflowlite_gpu_delegate.so*

2.3 Download the needed assets

$ cd ~/work/android_tflite
$ ./download_all_assets.sh

2.4 Build Android Applications

• Download and install Android Studio.
• Start Android Studio.

$ cd ${ANDROID_STUDIO_INSTALL_DIR}/android-studio/bin/
$ ./studio.sh

• Install NDK 20.0 by SDK Manager of Android Studio.
• Open application folder (eg. ~/work/android_tflite/tflite_posenet).
• Build and Run.

3. Tested Environment

Host PC	Target Device
x86_64	arm64-v8a
Ubuntu 18.04.4 LTS	Android 9 (API Level 28)
Android NDK r20b

4. Related Articles

• Android NDK 環境で TensorFlow Lite GPU Delegate を使う方法 (Qiita)

5. Acknowledgements

• https://github.com/google/mediapipe
• https://github.com/yu4u/age-gender-estimation
• https://github.com/TachibanaYoshino/AnimeGANv2
• https://github.com/openvinotoolkit/open_model_zoo/tree/master/demos/python_demos/human_pose_estimation_3d_demo
• https://github.com/ialhashim/DenseDepth
• https://github.com/MaybeShewill-CV/bisenetv2-tensorflow
• https://github.com/margaretmz/Selfie2Anime-with-TFLite
• https://github.com/NathanUA/U-2-Net
• https://tfhub.dev/sayakpaul/lite-model/east-text-detector/int8/1
• https://github.com/PINTO0309/PINTO_model_zoo