Face Recognition on STM32mp2

[20DeViL00]

Hi everyone,

I need some guidance on building a real-time face recognition pipeline on the STM32MP2 platform.

My requirements:

• Reliable face detection

• Good, consistent face embeddings

• Decent FPS on MP2

• Preferably ONNX or pure CPU, since heavy frameworks aren’t ideal

• Minimal external dependencies

• No reliance on TensorFlow or PyTorch

I’m looking for suggestions on:

• Which face detector works best on MP2 (YUNet / SCRFD / BlazeFace / SSD)?

• Which embedding model (ArcFace, MobileFaceNet, SFace, etc.) performs well on CPU?

• Any recommended ONNX + OpenCV DNN pipelines that others have successfully run on MP2?

• Tips for optimizing inference speed on this hardware

If anyone has experience or working examples, it would really help me understand the right direction to build a practical, efficient solution.

Thanks in advance!

[PatrickF]

Hi @20DeViL00 

please have a look to https://wiki.st.com/stm32mpu/wiki/Face_recognition

Regards,

[20DeViL00]

Hi @PatrickF ,

Thanks for your reply.

I’ve already tested the inbuilt face-recognition package that comes with the AI image, but unfortunately the results aren’t usable in my case. The overall accuracy is very low, and I’m also seeing frequent false detections during real-time testing.

Is there any recommended approach, updated model, or configuration from ST that can improve the performance?
If not, I’d really appreciate guidance on how to integrate a more reliable detection + recognition pipeline on the STM32MP2 platform.

Thanks!

[PatrickF]

Hi @20DeViL00 

On my side, I cannot help more (I'm more HW oriented).

ST sold SoC with Software examples (here, just speculating, the example is maybe not focusing recognition performance, but more inference time, ). 

I assume it is up to you to build you own AI model fitting your use cases (and aligned with the available NPU/CPU performance) using other tools and the open source SW we provide.

You could also have a look for potential help from ST partners : https://www.st.com/content/st_com/en/partner/partner-program.html 

Regards.