Multiple camera input to Deepstream through python for Object detection using SSD Mobilnet V2 on Jetson Nano

Hi everyone,

I am trying to implement python app which creates deepstream pipeline for object detection using SSD Mobilenet V2.
I tried this python file


It worked for single camera input. But it is too slow, lot of frames are getting lost from buffer which causes fps drop. Why this is happening?

And I want to give 2 camera inputs to Object detection model through deepstream python. What’s the process to do it?

Thanks.

Hi
You could follow this trouble shoot to tune the performance.
https://docs.nvidia.com/metropolis/deepstream/plugin-manual/index.html#page/DeepStream%20Plugins%20Development%20Guide/deepstream_plugin_troubleshooting.html#
Part: The DeepStream application is running slowly.

Thank you @amycao for reply.

I tried steps given on that link, still no improvement.
I used GETFPS() to get fps, it showed average 8 fps for both camera inputs but it is displaying output very slowly, maybe it is dropping frames before display.

And when I use config file for running the two csi camera input streams instead of python I got upto 17 fps for each camera stream, but for python it is dropping to half for each stream.

I need to trigger GPIO pins when desired object is detected. So I need to use python code for inference.

Do you mean the perf of deepstream c/c++ app is better than deepstream python apps, which app are u using?

Using deepstream-app command line app.

You could narrow down the difference between deepstream-app and python sample you used to see which part caused the perf drop.

deepstream-app uses config files to run, I built my python code based on that config files.
But it is still giving perf drop.

You could add queue after streammux for each element.