Object Detection / Multilabeling with the eIQ Boards (MIMXRT1060)

matthiasroedder · ‎01-13-2021

Hello Everyone,

First of all: Happy New Year!

I'm using the MIMXRT1060 and I'm able to run Algorithms for simple Object Classifications on it. For example reading a picture of a digit (0-9) and classify it or classifying if there is a dog or a cat on a picture.

Now I'm wondering if it is possible to do Object Detection and Multilabeling on the MIMXRT1060 board (e.g. YOLO algorithms) ?

For example to detect different animals and their location on a single picture?

Does anyone know if it is possible or not with such a board and if not which board is necessary for tasks like this?

Thank you in advance

Best Regards

Matt

Alexis_A · ‎01-21-2021

Hello @matthiasroedder ,

Sorry, I think I add the link in my previous message. This is the app note.

https://www.nxp.com/docs/en/user-guide/IMX-MACHINE-LEARNING-UG.pdf

Best Regards,

Alexis Andalon

View solution in original post

Alexis_A · ‎01-18-2021

Hello @matthiasroedder,

As far as I could check this kind of algorithm needs more processing the one you could find in this MCU. It could work but would be slow and maybe not practical, for this kind of algorithms and MPU would be more fitting to the application. The following app note mentions some MPU that supports these algorithms and some reference material.

Best Regards,
Alexis Andalon

carstengroen · ‎01-22-2021

I don't understand how you can conclude that this application demands more processing power that can be found in a RT1060 and that an "MPU" is needed ?

It all depends on how often one would need to classify an object, I can easily run TensorFlow Lite on a LPC54628 and do classification, I can't do 100 image classifications pr second, as well as you can't run 1000 classifications on a "MPU".

It all comes down to requirements.

You can even do this stuff on an old ARM7 LPC2458 running 72 MHz if the speed is enough for your application.

So before answering such a question, I think it is important to ask a few questions the other way....

matthiasroedder · ‎01-22-2021

@carstengroen Thanks for your reply.

In my case I don't have to classify that often. One image per second is okay for example.

I can't find any example code for object detection on microcontrollers... I don't know how to implement it.

carstengroen · ‎01-22-2021

@matthiasroedder ,

you're welcome

I have tensorflow lite running (for now, just for testing) on a LPC54628 processor with external SDRAM. If I run the "micro speech recognition" sample from TensorFlow, I can detect a word every 250 mS.

The sample: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech

A good place to start is here: https://www.tensorflow.org/lite/microcontrollers

Now, I can't of course guarantee you that you can do 1 classification of your image pr second, this depends of course on the resolution of the image. However, I would think that a RT106x would come really close for "reasonable resolutions". I would suggest to try out on a EVK for the 106x family (there is a small camera with the board, and there are samples in the SDK for that board).