IMX219 + NPU real-time object detection on Zero 3W (experimental)

#41

Are you using OpenCV? Those Bounding Boxes look a lot like OpenCV

#42

I would check GPU usage to make sure it is not being software rendered somehow. You should see 5%-7% load ( plain X without desktop enviroment )

#43

Reviewing the code, i computed the Inference + the Postprocess time as the total of 34 ms, the remaining 20 ms was rendering the image, rects and text.

#44

Hey Avaf, I will send my modified code shortly, maybe there is something different? I only did what you told me to do

#45

Here is the code I used as-is

./ff-rknn -p nv12 -f v4l2 -s 1024x600 -i /dev/video0 -m ./car-license-yolov5n.rknn -x 1024 -y 600

Built with jellyfin-rga , jellyfin-mpp and Nyanmisaka FFMPEG

1 Like
#46

hello, i am new for zero 3W, could you please tell me how the build the right firmware to light imx219, so i can get the image? thank you!

#47

Hi!

How did you managed to change NPU frequency? I’ve search the forum for clues, but the only one I got is /sys/class/devfreq/fde40000.npu/userspace/set_freq, but my image (https://github.com/radxa-build/radxa-zero3/releases/download/b6/radxa-zero3_debian_bullseye_xfce_b6.img.xz) does not have this file and userspace folder at all. Could you please provide your method for changing NPU frequency? My current NPU freq value is too low:

root@radxa-zero3:/sys/class/devfreq/fde40000.npu# cat cur_freq
0
#48

Great work! Kudos,

I & @mosamlife are working on a similar project.

Let’s discuss about the project and collaborate. We can figure out a way to work togather.

Thanks,
Ankit

IMG_2229
IMG_2229.jpg1206×718 996 KB

#49

Hi @avaf ,

I am working on a similar project with @mosamlife

Let’s discuss about the project and collaborate. We can figure out a way to work together.

Thanks,
Ankit - axnkit@gmail.com

11%20PM
WhatsApp Image 2025-01-23 at 12.50.11 PM.jpeg1198×1600 136 KB

#50

Can you share a bit more about your project?

Mendez and his friend have adapted ff-rknn sdl2 to his hardware and his project is mature and very advanced, you can adjust the code for your project.

It is shared here: https://github.com/PuntoMaximo/ff-rknn-sdl2

Today I am researching to build a quadcopter including a flight computer for learning purposes. I intend to share my experience if I finish it, and if it actually flies. So, learning by trial and error from zero experience in the field.

#51

We are trying Simultaneously perform…

  • Realtime Face/Object Detection using IMX219 using NPU of RK3566

  • Although Zero 3W gives 1080 output, we are playing 4K video content playback on HDMI using GPU

  • Running an application layer to communicate with cloud and download 4K content using CPU

And we’re struggling. Probably we’re doing some mistakes and we’re also lagging experience working on RK Chips. Can we collaborate? If you find it challenging, can you ping on axnkit@gmail.com