Object detection with MCIMX8QM

nagashima — Tue, 11 Aug 2020 23:58:29 GMT

Dear All,

MCIMX8QM-CPU: I am trying to detect motion using i.MX 8QuadMax, but the recognition is very slow.
The following URL is for reference.
https://www.nxp.com/docs/en/nxp/user-guides/UM11226.pdf

It takes about 40 seconds to recognize dogs, bicycles and cars in dog416.png in the PDF.
./example_dnn_object_detection -width=1024 -height=1024 -scale=0.00392 -rgb -zoo=models.yml yolo

Since the image size is 416x416 dots, set width and height to 416,
./example_dnn_object_detection -width=416 -height=416 -scale=0.00392 -rgb -zoo=models.yml yolo
And it was accelerated up to about 10 seconds.

Also, when it comes to recognition with camera images, the image is updated for about 4 minutes until the first display, and then only every few minutes thereafter.

Is this the limit for i.MX8QM?
Also, couldn't we use another method to increase the recognition frequency to around 5fps?

Best Regards

Re: Object detection with MCIMX8QM

Yuri — Wed, 12 Aug 2020 09:50:42 GMT

nagashima

Hello,

We do not have i.MX8QM performance estimations for eIQ applications.

In any case hardware acceleration improves situation.

Customers can refer to Coral Dev Board TPU | NXP

Also i.MX8MP is more suitable for machine learning and vision.

Regards,

Yuri.

i.MX ProcessorsのトピックRe: Object detection with MCIMX8QM

Object detection with MCIMX8QM

Re: Object detection with MCIMX8QM