Accelerating the MediaPipe models with Hailo

Hello RPI5 Hailo AI Kit community :slight_smile:

I would to share a project where I accelerated the MediaPipe models for Hailo-8:
Accelerating the MediaPipe models with Hailo-8

I am waiting for my RPI5 + Hailo-8L kit to test this out on the RPI5.

In the meantime, if anyone would like to try this out on their RPI5 Hailo AI Kit, please let me know …

I will only be supporting the “official” version of DFC supported by Raspberry Pi, which I believe is v3.27.0.

Enjoy !

1 Like

Fantastic work! Your article is incredibly detailed and well-explained.

About the effort to reduce the number of contexts, another way to address this is by enabling performance mode. This can be done with:

performance_param(compiler_optimization_level=max)

This will activate an algorithm that optimizes the utilization of the device’s various resources to their maximum potential.

1 Like

Thank you very much for those kind works, and for this valuable information.
I will definitely give this a try.

Regarding my “incredibly detailed and well-explained” article, there is still so much information missing …

I have a question related to the hand_landmark_lite model in the Hailo model zoo.

HAILO8_hand_landmark_detection.rst at master · hailo-ai/hailo_model_zoo · GitHub

It is the same model that I also target in my article, which I call version 0.10 lite of the hand landmarks model.

Do you know which dataset was used to quantize/compile this model (how did Hailo solve this issue of) ?
How is Hailo measuring the accuracy for the hand_landmark_lite model in the Hailo model zoo ?

My compiled model benchmarks at 3132 FPS (probably because of the 60% 4-bit setting, but possibly at the cost of lower accuracy), compared to the Hailo model zoo version which benchmarks at 2164 FPS.

Cheers !

Mario.

You can find the link for downloading the dataset here. That same script can be used for creating the tfrecord needed for the evaluation.

For measuring accuracy, you can use the hailomz eval tool.

1 Like

Thank you so much for this valuable feedback.

I also tried the “performance_param(compiler_optimization_level=max)” setting, and it works as expected, thank you :smile:

Every time I scratch deeper or try something new with the Hailo collateral, I am pleasantly impressed :sunglasses: :+1:

Very impressed with the Hailo solution, just WOW !

Mario (another happy user)

2 Likes