AM68A: Request for Customized Development with TI Boards

ASIF PATANKAR

Part Number: AM68A

Hi! 

My company bought a TI EVM with AM68A Processor. Going through your documentation I was able to rectify the development pipeline for segmentation cases. 
But feeling incomplete in the case of Keypoint detection. 
As the project demands more of just human pose detection which I understand TI EVM is already equipped with it. 
Can you give me references and suggestions where should I look for development of Keypoint Detection using models suggested in the SDK of TI?

over 2 years ago

0 Pratik Kedar over 2 years ago

TI__Mastermind 24041 points

Hi,

As I comprehend from your posted question, you are looking for road map to get started with custom model compilation(Keypoint Det).

You start with edgeai tidl tools repos, specifically OSRT Python examples here : https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master

We recommend you to study sample models from here and then start with custom model compilation journey.

Pratik

0 Manu Mathew over 2 years ago in reply to Pratik Kedar

TI__Genius 11486 points

What exactly do you mean by development of Keypoint detection models? Do you mean model training part of the model compilation part?

Notes:

====

We have examples of model compilation in edgeai-tidl-tools described above. Specifically for Keypoint detection, we have examples in https://github.com/TexasInstruments/edgeai-modelzoo and https://github.com/TexasInstruments/edgeai-benchmark

We support training of efficient Keypoint detection models using https://github.com/TexasInstruments/edgeai-yolox (YoloX-pose) and this method is described in the paper https://arxiv.org/abs/2204.06806 This method also have been incorporated into the popular mmpose repository: github.com/.../yolox_pose

And currently we are in the process of adding Keypoint detection flow using edgeai-yolox into https://github.com/TexasInstruments/edgeai-modelmaker

0 ASIF PATANKAR over 2 years ago in reply to Manu Mathew

Prodigy 20 points

Hi!

Thankful for your valuable response!

The fact that the segmentation required a little tweaking and the complete pipeline was sensible using the Model Maker. Whereas, the Keypoint Detection is has a multi-model compilation if using Model Maker and it doesn't seems to be complete if I want to train->validate->quantize->compile or vice versa. I will be eagerly waiting for workflow integrating MMPose into Model Maker.

Sincerely,
Asif

0 Manu Mathew over 2 years ago in reply to ASIF PATANKAR

TI__Genius 11486 points

edgeai-modelmaker and edgeai-yolox have been updated - can you pull both of these repositories and try Keypoint Detection once again?

If you get any error - please share the error here so that we can fix it. Thanks.

0 Manu Mathew over 2 years ago in reply to Manu Mathew

TI__Genius 11486 points

keypoint detection is currently hidden under a flag and not enabled by default until it tested more. In edgeai-modelmaker, set this flag to 1 and run the setup_all.sh to enable keypoint detection feature: https://github.com/TexasInstruments/edgeai-modelmaker/blob/master/setup_all.sh#L39

PLUGINS_ENABLE_EXTRA=1

0 ASIF PATANKAR over 2 years ago in reply to Manu Mathew

Prodigy 20 points

Hi!

Surely, I'll update you with the proceedings in the keypoint development. Presently, I'm testing the segmentation pipeline with some custom project dataset.

Thankful to you for the follow-up!

0 Pratik Kedar over 2 years ago in reply to ASIF PATANKAR

TI__Mastermind 24041 points

We recommend to create new thread if you face any issue with your experiments.

Closing the thread, thanks you.

Pratik

Processors

Processors forum

AM68A: Request for Customized Development with TI Boards