I’m going to deploy 3 object detection models on jetson xavier, my client said that each one model should be able to inference 5 web camera sources at the same time, so that means we should deal with 15 sources in total.
I know deepstream have some demo to show how to do the multiple camera sources inference, I’m not so sure if xavier has enough compute sources to do that. My models are basically ssd, or yolo3, with mobilenet backbone.
Meanwhile, we don’t need to run every model in real time, 5 FPS is ok for us, can I modify the deepstream demo to control the FPS? So that we could save some compute sources.
Having said that, we can also batch the frames from multiple sources, and input into triton which deployed ssd. Compare with deepstream solution, which one you guys recommend?
Thanks!
Hi
Deepstream 5 has a yolo plugin example.
Also you can refer this link.
5FPS * 15 source is not so much. You can achieve it.
Also Nvidia transfer learning tool can help you to speed up yolo detection.
You can search and try it as well.
Hi,
Please try DeepStream SDK, as jun suggested. After installation through SDKManager, you can see the packages in
/opt/nvidia/deepstream/
Is there any performance difference between python verson and c++ verson for deepstream demo? I noticed deepstream5.0 support python binding.
Hi,
When using Python to just construct the pipeline, the pipeline performance is comparable to the C/C++ version. When additional processing is done in the Python probe functions, some performance degradation is expected. The more processing done in Python, the worse the degradation. With light processing in Python, most pipelines can still run in real-time for some number of streams.