AM62A7: Face detection on AM62Ax: recommended models/frameworks (BlazeFace / MTCNN support)

Jardel Vieira

Part Number: AM62A7

Tool/software:

Hello TI team,

We are currently developing a system to recognize face using the AM62Ax SoC. One of our core features is real-time face detection, and we would like to understand which models or frameworks are recommended or officially supported by TI for this task.

Specifically:

Does TI provide any pre-integrated face detection models or frameworks in the Vision Apps or TIDL toolchain?
Is there official or community support for models like BlazeFace (MediaPipe) or MTCNN?
Are there known examples or documentation for running these models via ONNX or TFLite formats using TIDL?
Is there any face detection model optimized or validated for the AM62Ax platform?

Our goal is to deploy a lightweight model that works efficiently with the AM62Ax's CPU/NPU/DSP architecture, preferably using the existing TIDL + GStreamer pipelines.

Any guidance, links, or examples would be highly appreciated.

Thanks in advance!

2 months ago

0 Qutaiba Saleh 2 months ago

TI__Expert 4740 points

Hi Jardel,

It is our pleasure to help. In the edgeai-modelzoo, we have a couple of object detection models (yolox-s-lite and yolox-tiny-lite) which are trained for face detection using the widerface dataset. The artifacts for these models are provided at: https://github.com/TexasInstruments/edgeai-tensorlab/tree/r10.0/edgeai-modelzoo/modelartifacts/AM62A/8bits. For performance data bout these two models and other models look at the Model Selection tool in the Edge AI Studio: https://dev.ti.com/edgeaistudio/.

The yolox-s-lite is used in a demo for video conferencing. The demo combined both vision and audio processing. For your use case, you might want to ignore the audio part if you don't need it. The demo can be access at: https://dev.ti.com/tirex/explore/node?node=A__AEz622OMkJJnmUkmRkekmQ__EDGE-AI-DEMOS__PIzCUcf__LATEST and https://github.com/TexasInstruments-Sandbox/edgeai-demo-audio-visual.

Please, let us know if further details are needed.

Best regards,

Qutaiba

0 Jardel Vieira 2 months ago in reply to Qutaiba Saleh

Prodigy 10 points

Hello Qutaiba,
How are you?

Thank you very much for the helpful information on face detection. The YOLOX-Tiny-Lite and YOLOX-S-Lite models trained on WiderFace seem to be excellent candidates for our use case.

As a next step, we are looking into the face recognition part — specifically the embedding extractor, which will be used to compare against pre-registered driver profiles (similar to Dlib’s face_recognition library, which outputs 128D embeddings and stores them in a .pkl format).

We would like to know:

Are there any recommended models supported by the AM62Ax or the TIDL toolchain for face embedding extraction, such as ArcFace, MobileFaceNet, or other lightweight networks?
Are there any examples, references, or performance benchmarks for running these types of models on the AM62Ax?
Can these models be deployed through ONNX or TFLite formats with Edge AI Studio?

The goal is to combine the face detection model (e.g., YOLOX-Tiny-Lite) with an efficient face embedding model to enable recognition in-vehicle, with minimal latency and good accuracy.

Any guidance or suggestions are greatly appreciated.

Best regards,
Jardel

0 Qutaiba Saleh 2 months ago in reply to Jardel Vieira

TI__Expert 4740 points

Hi Jardel,

I am doing well. Thanks for asking. How about you?

We currently don't have a face recognition model ported to work on our AI accelerators. Our AI research team is considering this. That said, there are examples from third parties who have models which can work on ARM cores (A53 not HW AI accelerator) such this model which you can try online: https://website-git-automatic-ffid-plumerai.vercel.app/automatic-face-identification-live?max_framerate=12.

Some of the models available in the industry which you can look at: FaceNet512 for face embedding. This model might be easier to start with and it seems that most of the layers are supported by our TIDL. You can also look at cnn-facial-landmark which is built using Tensorflow.

At high level evaluation these models seems to be feasible to port using our edgeai-tidl tool: https://github.com/TexasInstruments/edgeai-tidl-tools. Examples of how to start with this tool are available at: https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master/examples/osrt_python. To evaluate a feasibility of compiling/porting a specific model to work on our HW AI Accelerators, you can refer to the list of supported layers at: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/supported_ops_rts_versions.md. Models with layers not included in the list can still be ported by implementing the unsupported layers on the ARM cores.

Please, let us know if any help is needed with the TIDL tool.

Best regards,

Qutaiba

Processors

Processors forum

AM62A7: Face detection on AM62Ax: recommended models/frameworks (BlazeFace / MTCNN support)