Request for Face Recognition Example Using NPU on i.MX8M Plus on Android Platform

[mohamed_gaseen]

Hi NXP Team,

I am currently working with the i.MX8M Plus board and leveraging its NPU accelerator for machine learning workloads.

At present, my requirement is to implement face recognition with NPU acceleration on the Android platform (not Yocto-based). I would appreciate it if you could share a working reference or sample code that demonstrates face recognition using the NPU on Android.

If there are any specific models, Android SDKs, or dependencies required, kindly include those details as well.

Looking forward to your guidance and support.

[Bio_TICFSL]

Hello,

You can see the face recognition demo on section 8.1.4 of the i.MX Machine Learning User's  guide, it works for android and yocto.

Regards

[mohamed_gaseen]

Hi @Bio_TICFSL,

Thank you for your timely response and support.

Regarding the Machine Learning User Guide link you provided — I’d like to clarify that we are specifically looking for Android-based examples rather than Python implementations.

For reference, we successfully built and tested the Image Classification Android example from the following GitHub repository:
https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/android

This approach worked well for us.

We are now looking to implement Face Recognition using a similar Android-native approach. Would it be possible for you to share an Android example or source code for face recognition, ideally in a structure similar to the image classification example?

Having access to such an implementation would be extremely helpful for our ongoing work and integration.

Thanks in Advance.

[Bio_TICFSL]

Hello,

The demo is working in Android BSP, but from the console I think what you are looking is an app.

Regards

[mohamed_gaseen]

Hi,

The default application (android.example.com.tflitecamerademo) runs successfully on the Android 14 BSP and primarily performs image classification. However, we require a face recognition demo application to run on the same BSP.

Could you please confirm whether NNAPI is equivalent to the NPU?

We observed a latency of 200 ms when using the CPU, whereas NNAPI achieves 12 ms.

[Bio_TICFSL]

Hello,

NNAPI is Neural Network API de android, This API use components of GPU and NPU.

Regards

[mohamed_gaseen]

Thank you @Bio_TICFSL for your prompt response and for confirming NNAPI support on the i.MX8MP platform.

As a next step, I am looking for a face recognition example implemented in Android code, preferably utilizing NNAPI delegates to run on the NPU. While I have reviewed the Python-based examples provided in the eIQ documentation, I would appreciate a native Android application—similar to TensorFlow Lite Android examples—but tailored for face recognition.

Also could you please reply to these two points:

• Do you have any VIT voice recognition running model on Android ?
• Can you please find the attached image, here we're selecting NNAPI which is both GPU & NPU supported as you referred. But I want to specify only for GPU. Is there any possible way ?

Kindly share a reference implementation or point me to any existing Android project that demonstrates face detection or recognition on i.MX8MP using NNAPI?

[Bio_TICFSL]

Hi,

Yes, you can apply this:

https://github.com/nxp-imx-support/imx-voiceplayer

Regards

[mohamed_gaseen]

Hi @Bio_TICFSL ,

Is there any update regarding this??

Please support me with it ASAP.

[varshilg]

Hi @mohamed_gaseen ,

Glad to know that you are working on the IMX8M Plus board with NPU. I just want to clarify: since you mentioned that you are working on the Android platform, does that mean you are deploying your model not on a Yocto-based source, but rather using the AOSP source for deployment?

Regards,
Varshil

[mohamed_gaseen]

Dear @varshilg,

Yes we need in the AOSP source for the face recognition, not in the Yocto.

[varshilg]

Okay, great.
Are you only using pre-trained models, or are you building your own custom models from scratch? I’m asking because I am also deploying a model on the IMX and facing an issue where the Fully_Connected layers are not being converted to a Neuron Graph for the NPU.