[QUESTION] Scale Operations with Multiple Videos

[pedromoraesh]

I'm currently working on a job where we need to process 400 Hours of video every day in a RTX4090. I was able to decode 1 hour of video in 30seconds using pynvvideocodec and CV-CUDA.

Using the provided sample of object detection with PeopleNet from Nvidia i was able to run 1 hour in 3 minutes. But now i need to scale it in my cloud. I have a few questions:

• Is worth spawn subprocess to decode video and inference using the same context or multiple context?
• What is safest approach: run multiple python programs to avoid concurrency or run the same python process with threading/multiprocess (i don't think is a good a idea dude do GPU handling),?

I'm trying to avoid as much as possible trade between CPU and GPU. My next step is track those people using ByteTrack wich i will try to port to GPU entirely.

Tips for those who are trying to run the sample:

• Need TensorRT 8.6.x tar file from NVIDIA
• tao-converter can be downloaded, but is deprecated, you also need to chmod +x tao-converter and only works with tensorrt 8.x.x.
• Also install python whl from TensorRT tar file.