Need help running object detection with MediaPipe Face Landmark on STM32N6570-DK

[Rutva Patel]

Hi ST Community,

I am working on the STM32N6570-DK board and need guidance on running two AI models together in a single pipeline:

1. YOLO-FastestV2 for object detection (this is my custom train model)
2. MediaPipe Face Landmark for facial landmark detection

My Goal

I want to:

• Run YOLO-FastestV2 to detect objects/faces from camera input
• Then run MediaPipe Face Landmark on detected face regions (ROI)
• Process both models efficiently on STM32N6570-DK
• Display final output with bounding boxes + face landmarks

Current Status

• I can run object detection examples on STM32N6570-DK.
• I have AI models available in ONNX/TFLite format.
• I am exploring deployment using STM32Cube.AI / ST AI tools.
• I am not sure about the best architecture for running both models together.

Questions

I need help with the following:

1. Is it possible to run two AI models in parallel on STM32N6570-DK?
2. What is the recommended pipeline?
   • Camera → YOLO-FastestV2 → ROI extraction → Face Landmark → Display Output
3. Should both models be merged into one application, or should I create separate inference pipelines?
4. Are there memory limitations when using both models together?
5. Is there any example project for multi-model inference on STM32N6?
6. What is the best way to handle ROI cropping/resizing for the second model?

Challenges

Main concerns are:

• RAM/Flash usage
• Inference latency
• Camera pipeline integration
• Efficient post-processing

If anyone has experience with multi-model deployment or similar use cases on STM32N6 devices, your suggestions would be very helpful.

Thanks in advance.

[Julian E.]

Hi ​@Rutva Patel,

 

Please look at this example:

STMicroelectronics/x-cube-n6-ai-hand-landmarks: An AI software application package demonstrating hand landmarks detection use case on STM32N6 product.

 

We use 2 models: hand detection, then hand landmark with some ROI between the 2 models.

I would suggest giving it to claude code or any good LLM and it should be able to answer any of your questions

 

the flow is as follow:

 

1. The output of the “Palm detector” are multiple bounding boxes. We select one and this is a small bounding box (in red in the demo).
   We are then resizing it and rotating the image to have the hand in this position ✋(this kind of is the green bounding box)
   This is then sent to the other model
2. The output of the “Hand landmark model” is both the landmarks and if a hand is detected
   if the confidence of a hand being in the image too low => we go back to the hand detector (“No” path)
   Else, we get a similar bounding box as the one mentioned in point 1 (“Yes” path)
3. We apply the same preprocessing (because the hand may have moved a little) => resize + rotation before feeding it again in the “hand landmark” model

 

We do not support epoch level parallelization (run 1 epoch of model1, then run 1 epoch of model2). Maybe with a RTOS you could do it.

For the best way to handle ROI, as I said ask a LLM to see what is done in the package. This is also described here:
x-cube-n6-ai-hand-landmarks/Doc/Application-Overview.md at main · STMicroelectronics/x-cube-n6-ai-hand-landmarks

 

Have a good day,

Julian