I am running detection algorithm on jetson nano and initial frames (2-5) are taking too much time to process or stuck and it creates a bad user experience every time we run any algorithms. Does anybody know how to resolve or improve this
Do you use OpenCV to process camera?
If yes, have you enabled GStreamer support.
I use opencv that comes with the jetson nano image. I don’t think it is opencv issue, because it happens only when i pass the first few frames to tensorflow session it is getting stuck or taking longer time to do the prediction.
Does your nano work at MAX-N performance mode?
sudo jetson_clocks or sudo nvpmodel -m 0
In dynamic mode. the frequency is adjust according to the workload. Maybe need some time to speed up.
You can try to see if it works.
I tried those commands already and it does not resolve the issue. I have observed the same issue even when i am using tensorrt. Processing time of first few frames are very slow
TensorFlow initialization time is pretty long.
How do you know the issue comes from the camera rather than TensorFlow?
Would you mind to profile your application with Nsight and share the log with us first?