Issues with performance using PVA and CUDA Backends with VPI Library on Jetson Orin

[paarth.desai]

Hello, I am having performance issues when using the harris corner detector and OpticalFlowPyrLK functions in the VPI Library on the Jetson Orin with the PVA and CUDA backends. Specifically, when I use these functions with the named backends(PVA for harris corner detection and CUDA for optical flow), I see that the feature tracking done by the optical flow function has a severe drop off after a short period of time, regardless of where I begin my video. I also believe that the corner detection is worse with the PVA backend. For context, when I use the CPU backend, both of these functions work as expected with good performance, but I would like to be using the backends which I stated. Please let me know if you have any way to resolve this issue, thanks!

[AastaLLL]

Hi,

For the performance, please check below document to boost the device:
https://docs.nvidia.com/vpi/algo_performance.html

For the corner case issue, would you mind sharing an sample image pair or video to reproduce the difference between CPU and PVA backend?
We need this to check with our internal team for more information.

Thanks.