Hi,
This is expected since OpenCV only accepts CPU buffer data in BGR format. This would need to take significant CPU usage. If your use-case uses CUDA filters in OpenCV, please check this:
Nano not using GPU with gstreamer/python. Slow FPS, dropped frames - #8 by DaneLLL
It is to run a gstreamer pipeline and map NvBuffer to cv:;gpu::gpuMat. This eliminate additional memory copy. Or may check if the required functions are supported in VPI:
VPI - Vision Programming Interface: Main Page
So that you can switch to use VPI.