Improve the arithmetic performance of the OAK-1 processor

Jalen_Xu

Hi Luxonis Support Team,

We have now deployed three yolov5s neural network models on the OAK-1 camera, but we find that it runs very slowly. It takes a few seconds to process a frame. Is there any way to improve the computational efficiency? Or is there a better solution to improve the arithmetic performance of the processor?

I'm looking forward to your reply. Thank you very much!

erik

Hi Jalen_Xu ,
What is the input (frame) size of your yolo5s? We have benchmark results here about the RVC2 NN performance: https://docs.luxonis.com/projects/hardware/en/latest/pages/rvc/rvc2/#rvc2-nn-performance
If you have large input size (eg. 1MP), then slow FPS is expected, and the only way to increase it would be to use lighter architecture (eg. YOLO nano/tiny instead of small), or have smaller input size.

Jalen_Xu

erik Thanks for your reply! The input frame size is 460x460. I can not find the yolov5s performance in the link. Another thing is that we will deploy neural network models on a camera with a number of 3.

erik

Hi Jalen_Xu,
I apologize for the late reply. We don't have performance for yolov5s specifically, but we do have it for quite a several variations of yolov6 (different sizes and architectures), from which you could estimate the max speed of yolov5s.

network models on a camera with a number of 3.

Could you elaborate on this?

Jalen_Xu

Hi Erik,

Thanks for your reply!

In one project we used three yolov5 neural network models simultaneously. One for detecting the vehicle and then intercepting the image of the vehicle target from the original image, one for detecting the license plate region, and finally one for OCR to recognize the license plate numbers and letters.