General inquires CPU/ NvTRacker

I have the below general questions

  1. I’m using facedetect as primary GIE nv-infer. It takes input of 436x712 however i was advised to change nvstreammux dimensions based on input source dimensions, and it improved the FPS. I was wondering is it better to change nvstreammux dimensions to facedetect dimensions to avoid preprocessing in the GPU? as i don’t get why i have to preprocess the size of the input frame twice.
    And is it better to process the size in the cpu or gpu?
  2. what is the main role of the CPU in the device ‘jetson agx orin’ i noticed when enabling DLA, CPU load increase compared to the case when GPU is enabled only. Although the CPU doesn’t handle role in transferring data since the memory can be accessed by all hardwares ‘DRAM’.
  3. Does NvDCF tracker uses any deep learning model to extract features for Visual appearance similarity?

We perfer to use GPU instead of CPU for video process. Here is the guide for nvtracker: Gst-nvtracker — DeepStream documentation 6.4 documentation