Qengineering
Question about FPS data
Hii,
As usual thank you so much for what you are doing for the community!
I wanted to know if the FPS values shown in the table are for model inference on NPU or with ncnn?
I have run custom yolov5s models on the orangepi5 using npu but it only gives me upto 25FPS. Since both are RK3588, and theres very little performance difference between RK3588 and 3588s, I want to know what can be done to get my FPS to 37 and above.
Thanks in advance!
The figures reflecting the NPU. I've have measured them accurately. One thing I can think of is the adaption of YoloV5 for the NPU architecture. see: https://github.com/Qengineering/rknn_model_zoo/tree/main/examples/yolov5/model_comparison
Does your FPS calculation include the post_ processing time or is it just model processing thats calculated here?
It includes PostProcessing also. Look at the source code: https://github.com/Qengineering/YoloV5-NPU/blob/main/src/main.cpp In line 192, the start time is measured. Until then, we only have captured the frame. At line 266, the end time is set. Everything is done for now, except the imshow() on the screen.
Some time ago I got an email from Anton Maltsev (https://www.youtube.com/watch?v=e_BtAdS5M70) experiencing something similar.
Hi, Rients! Thank you very much! I will try to do this in a few weeks. I am really confused about the problem with my Radxa Zero 3W. I saw this once with Orange Pi5, but the problem was with drivers installation.
Perhaps a clue?
I see Yeah it could be driver and toolkit version differences i assume that could be the only reason. I'll try changing the latest librknnrt.so file and checking. Which version of rknn-toolkit are you using? I was working with 1.6.0 so I assume that couldnt be the reason.
The latest found on GitHub version 1.6.0. https://github.com/airockchip/rknn-toolkit2/