Not working on Android about tf-lite-unity-sample

[kaiquesacchi]

Yesterday we found out about this config and solved the problem in that exact way (the first option you suggested)!
Anyway, thank you so much for taking your time to help us. Your solution works. :) @ganaware

I have had a similer problem.
The cause of the problem is that Mono is the default scripting backend for a new unity project for Android.
Unfortunately, Mono can run only for the ARMv7 architecture.

So my solution is that:

1. Project Settings > Player > Android > Other Settings > Configuration

   • Select IL2CPP for Scripting Backend
   • Uncheck ARMv7 and check ARM64 for Target Architectures

2. Or, build my own TFLite library for ARMv7

   • libnnapi_delegate.so, libtensorflowlite_c.so, libtensorflowlite_gpu_delegate.so
   • They worked fine for me.

[ganaware]

I have had a similer problem.
The cause of the problem is that Mono is the default scripting backend for a new unity project for Android.
Unfortunately, Mono can run only for the ARMv7 architecture.

So my solution is that:

1. Project Settings > Player > Android > Other Settings > Configuration
   • Select IL2CPP for Scripting Backend
   • Uncheck ARMv7 and check ARM64 for Target Architectures
2. Or, build my own TFLite library for ARMv7
   • libnnapi_delegate.so, libtensorflowlite_c.so, libtensorflowlite_gpu_delegate.so
   • They worked fine for me.

[FrankSpalteholz]

Hi Roberto, i was facing similar issues on iOS with my iPad. In my case it had to do with the orientation (landscape, portrait etc). So most front-camera-sensors in smart-phones and tablets are rotated by 90 degrees and although Unity is able to compensate for that the inputTexture (WebCamTexture) is not taking account for that. So i had to flip the xAxes of the keypoints (1f-keypoint) and also flip uv's and normals of the face-mesh, but this only works on landscape-mode. You could also rotate your rendercam and mirror the canvas. In the end i was using OpenCV to adjust the textures how needed because it was the cleanest solution for my purposes. Hope this helps.

[asus4]

Thanks for sharing it @FrankSpalteholz

@Roberto7766 Do you get any errors on logcat?

[Roberto7766]

Thanks @FrankSpalteholz and @asus4.
Using logcat I found out that the issue is with the neural network. If I comment the NN part the camera works well.
Here the logcat result:

2020-10-30 08:09:02.815 20668-20687/? E/Unity: DllNotFoundException: libtensorflowlite_c
        at (wrapper managed-to-native) TensorFlowLite.InterpreterOptions.TfLiteInterpreterOptionsCreate()
      at TensorFlowLite.InterpreterOptions..ctor () [0x00006] in <9d167a8ec82e4b91b0e30378348ced48>:0
      at TensorFlowLite.BaseImagePredictor`1[T]..ctor (System.String modelPath, System.Boolean useGPU) [0x00006] in :0
      at TensorFlowLite.SSD..ctor (System.String modelPath) [0x00035] in :0
      at SsdSample.Start () [0x00011] in :0   

  (Filename: Line: 0)
2020-10-30 08:09:07.986 20668-20687/? E/Unity: NullReferenceException: Object reference not set to an instance of an object
      at SsdSample.Update () [0x00000] in :0 

Maybe I did something wrong setting up the project?

[asus4]

It seems like failing to load the TensorFlow Lite library. Please make sure that the CPU isn't 32bit. But I've confirmed some cases that 64bit Android fails as well.

[Roberto7766]

It is 64bit. What can be done to move forward? Thks.

[asus4]

The only thing I could see from the logcat is that the library is failing to load. We need to investigate the cause of the problem.

[kaiquesacchi]

I've cloned the repo and got it to work on Android, Windows and Linux. I've even added my own Image Classification model, created a new scene, loaded it and got it to work as well.

When I tried creating a new Unity project and then adding the package (using the Package Manager and selecting the package file given in the Packages folder of your repo), I only got it to work with Windows and Linux. For some reason, when using it on Android, it breaks after trying to create instances of InterpreterOptions and Interpreter.

Are you aware of any additional steps that should be performed in order to add the package to a new project?