tfliteSSDminimalWorkingExample

TensorFlow Lite SSD (Object Detection) Minimal Working Example for iOS and Android

Here are instructions for building and running a minimal working example of TensorFlow Lite SSD/object detection on iOS and Android.

How to build and run a minimal working example of TensorFlow Lite SSD/object detection on iOS

Part 1 (create a new iOS project with OpenCV libraries installed)

Note if using the sample project github.com/baxterai/tflitessdminimalworkingexample/ios;

• delete the existing references (red) to opencv2.framework in;
  • Project Navigator - Frameworks
  • Project Navigator - [APPLICATIONNAME] project (blue document icon) - [default target] - Build Phases - Link Binary With Libraries
• Download the opencv framework (https://opencv.org/releases.html , e.g. https://sourceforge.net/projects/opencvlibrary/files/opencv-ios/3.4.2/opencv-3.4.2-ios-framework.zip/download )
• Project Navigator - select [APPLICATIONNAME] project (blue document icon) - [default target]
• Build Phases - Link Binary With Libraries - Click the '+' - Add Other... - browse and select opencv2.framework (e.g. /Users/[USERNAME]/Documents/libraries/opencv2.framework)
• Build Settings - Framework Search Paths - add [INSERT COMPLETE PATH OF opencv2.framework] (e.g. /Users/[USERNAME]/Documents/libraries/)
• Project Navigator - select [APPLICATIONNAME] project (blue document icon) - [default target] - Build Phases

Otherwise, create an iOS Objective-C project with OpenCV (these remaining instructions can be skipped if using the sample project github.com/baxterai/tflitessdminimalworkingexample/ios):

[Extract from https://stackoverflow.com/questions/52030819/how-to-create-an-ios-objective-c-project-with-opencv/52030820#52030820 ] (instructions based on https://medium.com/@yiweini/opencv-with-swift-step-by-step-c3cc1d1ee5f1 )

• Download the opencv framework (https://opencv.org/releases.html , e.g. https://sourceforge.net/projects/opencvlibrary/files/opencv-ios/3.4.2/opencv-3.4.2-ios-framework.zip/download )
• Project Navigator - select [APPLICATIONNAME] project (blue document icon) - [default target]
• Build Phases - Link Binary With Libraries - Click the '+' - Add Other... - browse and select opencv2.framework (e.g. /Users/[USERNAME]/Documents/libraries/opencv2.framework)
• Build Phases - Link Binary With Libraries - Click the '+' - add these additional frameworks (for opencv); AssetsLibrary, CoreGraphics, CoreMedia, CoreFoundation, Accelerate, [UIKit, Foundation, CoreVideo, CoreImage]
• Build Settings - Framework Search Paths - add $(PROJECT_DIR)
• Build Settings - Framework Search Paths - add [INSERT COMPLETE PATH OF opencv2.framework] (e.g. /Users/[USERNAME]/Documents/libraries/)
• File - New - File - Cocoa Touch Class. Name it OpenCVWrapper and choose objective-C for Language.
• OpenCVWrapper.m and rename the file extension to .mm
• Manually change OpenCVWrapper.m to OpenCVWrapper.mm in the file header also
• Go to OpenCVWrapper.mm and add the following import statement on the top; #import <opencv2/opencv.hpp>

Part 2 (install the TFlite library)

Note if using the sample project github.com/baxterai/tflitessdminimalworkingexample/ios;

• delete the existing references (red) to libtensorflow-lite.a in;
  • Project Navigator - Frameworks
  • Project Navigator - [APPLICATIONNAME] project (blue document icon) - [default target] - Build Phases - Link Binary With Libraries
• build the libtensorflow-lite library and recreate the references to it (implement the first section of the following instructions);

[Extract from https://stackoverflow.com/questions/52030130/how-to-build-and-run-the-tensorflow-lite-ios-examples/52030131#52030131 ] (instructions based on https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/g3doc/ios.md )

• git clone https://github.com/tensorflow/tensorflow
• cd tensorflow
• git checkout r1.10 (select a version of tensorflow that contains tensorflow/contrib/lite/download_dependencies.sh)
  • [these instructions are based on https://github.com/tensorflow/tensorflow/tree/r1.10 ]
• cd tensorflow/contrib/lite
• ./download_dependencies.sh
• ./build_ios_universal_lib.sh
• cd examples/ios
• ./download_models.sh (download tensorflow models to simple/data and camera/data)
• Show the project navigator
• Update various settings in your app to link against TensorFlow Lite (see tensorflow/contrib/lite/examples/ios/simple/simple.xcodeproj / camera/tflite_camera_example.xcworkspace for example):
  • select [APPLICATIONNAME] project (blue document icon) - [default target] - Build Phases
  • Link Binary With Libraries - add (+) library tensorflow/contrib/lite/gen/lib/libtensorflow-lite.a
  • select [APPLICATIONNAME] project (blue document icon) - [default target] - Build Settings
  • Library Search Paths - add [INSERTTENSORFLOWSOURCELOCATIONHERE]/tensorflow/contrib/lite/gen/lib
  • Header Search paths - add [INSERTTENSORFLOWSOURCELOCATIONHERE] (the root folder of the tensorflow git repository)
  • Header Search paths - add [INSERTTENSORFLOWSOURCELOCATIONHERE]/tensorflow/contrib/lite/downloads
  • Header Search paths - add [INSERTTENSORFLOWSOURCELOCATIONHERE]/tensorflow/contrib/lite/downloads/flatbuffers/include
  • C++11 support (or later) should be enabled by setting C++ Language Dialect to GNU++11 (or GNU++14), and C++ Standard Library to libc++.

These remaining instructions can be skipped if using the sample project github.com/baxterai/tflitessdminimalworkingexample/ios;

• project navigator - in [INSERTAPPLICATIONNAME] project (blue document icon) - create new group called data
• drag and drop data folder items from tensorflow/contrib/lite/example/ios/simple/data (grace_hopper.jpg, labels.txt, mobilenet...) to the newly created data folder in xCode (when asked, select Destination: Copy items if needed)
• modify the application's existing AppDelegate.m accordingly with tensorflow/contrib/lite/examples/ios/simple/AppDelegate.mm/.h contents
• drag and drop source code items from tensorflow/contrib/lite/example/ios/simple (ios_image_load.h, ios_image_load.mm, RunModelViewController.h, RunModelViewController.mm, RunModelViewController.xib [NOT: AppDelegate.h, AppDelegate.mm, main.mm])

Part 3 (Create a patched version of obj_detect_lite.cc)

Note these instructions can be be skipped by using the obj_detect_lite-ios.cpp code in github.com/baxterai/tflitessdminimalworkingexample/ios;

Instructions based on; https://github.com/tensorflow/tensorflow/issues/15633 - hidden items - load more

• manually create a patched version of obj_detect_lite.cc
• git clone https://github.com/YijinLiu/tf-cpu
• OPTIONAL: git checkout master / 11b8a2c8f4fd882c591c05d9cc7fa4ed538e2661 (these instructions are based on the 14 August 2018 version, https://github.com/YijinLiu/tf-cpu/commit/11b8a2c8f4fd882c591c05d9cc7fa4ed538e2661)
• method 1;
  • add the reference nms(objects) along with the functions iou/objectComp/nms
  • add the struct Object defintition along with its references in class ObjDetector (std::vector<Object> objects;) and function AnnotateMat
• method 2;
  • copy the latest version of obj_detect_lite.cc by YijinLiu tf-cpu/benchmark/obj_detect_lite.cc to a temporary folder vMaster
  • git checkout f31fb25d97adcd00124646c0413bfdec989840fa (checkout the older 28 May 2018 version of https://github.com/YijinLiu/tf-cpu upon which WeiboXu's version was created, such that a diff can be generated)
  • copy the older 28 May 2018 version of tf-cpu/benchmark/obj_detect_lite.cc by YijinLiu to a temporary folder vOrig
  • copy the patched version of obj_detect_lite.cc by WeiboXu from https://github.com/tensorflow/tensorflow/files/2153563/obj_detect_lite.cc.zip to a temporary folder vOrigPatched
  • diff vOrig vOrigPatched (use windiff/kdiff3 for GUI)
  • note the changes to the file which must be integrated into the vMaster version of obj_detect_lite.cc
• Alternatively, download the TF Lite SSD minimal working example IOS source from github.com/baxterai/tflitessdminimalworkingexample/ios

Part 4 (Prepare the user interface)

Note these instructions can be be skipped using the UI code in github.com/baxterai/tflitessdminimalworkingexample/ios;

• Project Navigator (left pane) - [INSERTAPPLICATIONNAME]
• Open Main.storyboard
• Select View Controller Scene - View Controller - View
• Select the Library button (top right: circle with inside square)
• Search for 'Image View'
• Drag and drop a new View Image Object into View Controller Scene - View Controller - View (adjacent 'Safe Area')
• Create an outlet connection for the new Image View object
• Select the Assistant Editor button (top right: two intersecting circles)
• Select file ViewController.h in the Project Navigator
• If necessary hide the left and right panes to create more space (top right: blue box with left vertical bar, blue box with right vertical bar)
• hold the Ctrl key and drag the new image view object to the ViewController.h file (immediately below @interface ViewController: UIViewController) - name the object reference as imageViewInstance0

Part 6 (Add the TFLite SSD execution files)

Note these instructions can be be skipped using the project github.com/baxterai/tflitessdminimalworkingexample/ios;

• Download the TF Lite SSD minimal working example IOS source from github.com/baxterai/tflitessdminimalworkingexample/ios
• Project Navigator (left pane) - [APPLICATIONNAME]
• Drag and drop the relevant source code items from tflitessdminimalworkingexample (object_detect_lite-ios.cpp/object_detect_lite-ios.hpp/OpenCVWrapper.hpp/OpenCVWrapper.mm) into the primary application folder in xCode (when asked, select Destination: Copy items if needed)
• Merge the relevant code from tflitessdminimalworkingexample ViewController.m into [APPLICATIONNAME] ViewController.m
• Project navigator - in [APPLICATIONNAME] project (blue document icon) - create new group called data
• Drag and drop data folder items from tflitessdminimalworkingexample/data (grace_hopper.jpg, coco10_labels.txt, coco10.tflite) to the newly created data folder in xCode (when asked, select Destination: Copy items if needed)

How to build and run a minimal working example of TensorFlow Lite SSD/object detection on Android

Note these instructions can be be skipped using the project github.com/baxterai/tflitessdminimalworkingexample/android;

Instructions based on; https://stackoverflow.com/a/51969021/2585501 (https://stackoverflow.com/questions/50330184/build-and-run-tensorflow-lite-demo-with-gradle)

Note these instructions should theoretically support optional object tracking by adding the libtensorflow_demo.so library (although adding this library causes the app to crash; it may relate to a conflicting package naming scheme or gradle build conflict)

Part 1 (create a new Android project)

Android Studio - create a new Android project

• Basic
• Application name: e.g. [TF Lite SSD Minimal Working Example]
• Company domanin: e.g. [user.example.com]
• Project location: e.g. [.../tfliteSSDminimalWorkingExampleAndroid/]
• Include C++ Support
• generated package name: [com.example.user.tflitessdminimalworkingexample]
• Next
• Select the form factors and minimum SDK
• Phone and Tablet - e.g. [API 25: Android 7.1.1 (Nougat)]
• Next
• Add an Activity to Mobile
• Empty Activity
• Next
• Configure Activity
  • Activity Name: [MainActivity]
  • tick generate layout file
  • Layout Name: activity_name
  • Backwards compatibility (AppCompat)
  • Next
• Customise C++ Support
  • C++ Standard: Toolchain Default

Part 2 (prepare the Android project)

• Open the android project (tflitessdminimalworkingexample)
• edit app/build.gradle
• add the following code;

android {

...

    lintOptions {
        abortOnError false
    }    
    aaptOptions {
        noCompress "tflite"
    }

    compileOptions {
        sourceCompatibility JavaVersion.VERSION_1_8
        targetCompatibility JavaVersion.VERSION_1_8
    }
}

repositories {
    maven {
        url 'https://google.bintray.com/tensorflow'
    }
}

// import DownloadModels task
project.ext.ASSET_DIR = projectDir.toString() + '/src/main/assets'
project.ext.TMP_DIR   = project.buildDir.toString() + '/downloads'

// Download default models; if you wish to use your own models then
// place them in the "assets" directory and comment out this line.
apply from: "download-models.gradle"

dependencies {

...

    implementation 'org.tensorflow:tensorflow-lite:+'

Part 3 (copy the minimal TensorFlow Lite Android example source code)

• git clone https://github.com/tensorflow/tensorflow
• cd tensorflow
• OPTIONAL: git checkout master / 938a3b77797164db736a1006a7656326240baa59
  • [these instructions are based on https://github.com/tensorflow/tensorflow/commit/938a3b77797164db736a1006a7656326240baa59 ]
• open tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/ in file explorer
• copy AutoFitTextureView.java, Classifier.java, OverlayView.java, TFLiteObjectDetectionAPIModel.java to [PATHTOAPPLICATION]/app/src/main/java/com/example/user/[APPLICATIONNAME]
• copy AutoFitTextureView.java, Classifier.java, OverlayView.java, TFLiteObjectDetectionAPIModel.java to [PATHTOAPPLICATION]/app/src/main/java/com/example/user/[APPLICATIONNAME]
• copy env BorderedText.java, ImageUtils.java, Logger.java, Size.java to [PATHTOAPPLICATION]/app/src/main/java/com/example/user/[APPLICATIONNAME]/env
• copy env MultiBoxTracker.java, ObjectTracker.java to [PATHTOAPPLICATION]/app/src/main/java/com/example/user/tflitessdminimalworkingexample/tracking
• in all copied java files, replace package org.tensorflow.demo.env (e.g. with package com.example.user.[APPLICATIONNAME] where this matches the [PATHTOAPPLICATION]/app/src/main/java/com/example/user/[APPLICATIONNAME] directory structure)
• in all copied java files, replace all instances of com.example.user.tflitessdminimalworkingexample (e.g. with com.example.user.[APPLICATIONNAME] where this matches the [PATHTOAPPLICATION]/app/src/main/java/com/example/user/[APPLICATIONNAME] directory structure)

Part 4 (copy the minimal TensorFlow Lite Android example UI code)

• modify [PATHTOAPPLICATION]/app/src/main/res/activity_main.xml

• change android.support.constraint.ConstraintLayout to android.support.design.widget.CoordinatorLayout in [PATHTOAPPLICATION]/app/src/main/res/activity_main.xml;

• in 'Design' (or 'Text') view, add an ImageView object called imageView1 to the component tree - CoordinatorLayout

• open tensorflow/contrib/lite/examples/android/app/src/main/res/ in file explorer

• import the layoutcamera_connection_fragment_tracking.xml contents into [PATHTOAPPLICATION]/app/src/main/res/activity_main.xml;

  <?xml version="1.0" encoding="utf-8"?>
  <android.support.design.widget.CoordinatorLayout xmlns:android="http://schemas.android.com/apk/res/android"
      xmlns:app="http://schemas.android.com/apk/res-auto"
      xmlns:tools="http://schemas.android.com/tools"
      android:layout_width="match_parent"
      android:layout_height="match_parent"
      tools:context=".MainActivity">
  
      <ImageView
          android:id="@+id/imageView1"
          android:scaleType="fitStart"
          android:layout_width="match_parent"
          android:layout_height="wrap_content"
          app:srcCompat="@android:color/holo_green_dark" />
  
      <com.example.user.[APPLICATIONNAME].OverlayView
          android:id="@+id/tracking_overlay"
          android:layout_width="match_parent"
          android:layout_height="match_parent"/>
  
      <com.example.user.[APPLICATIONNAME].OverlayView
          android:id="@+id/debug_overlay"
          android:layout_width="match_parent"
          android:layout_height="match_parent"/>
  
  </android.support.design.widget.CoordinatorLayout>
  

Part 5 (create the TensorFlow Lite Android example UI minimal activity source code)

• Download the TF Lite SSD minimal working example IOS source from github.com/baxterai/tflitessdminimalworkingexample/android
• insert the tfliteSSDminimalWorkingExample/app/src/main/java/com/example/user/tflitessdminimalworkingexample/MainActivity.java contents into `[PATHTOAPPLICATION]/app/src/main/java/com/example/user/tflitessdminimalworkingexample/MainActivity.java
• Gradle Sync
• Build
• Run
• [give permissions to app when requested]
• Run example app (tflitessdminimalworkingexample) on Android phone (search - tflitessdminimalworkingexample)