Hi,
This is possible in OpenCV since there is additional memory copy. Please refer to discussion in
[Gstreamer] nvvidconv, BGR as INPUT
The performance would better if you can run pure gstreamer pipeline, or use cv::gpu::gpuMat as demonstrated in the sample:
Nano not using GPU with gstreamer/python. Slow FPS, dropped frames - #8 by DaneLLL
If you must run gstreamer pipeline in cv2.VideoCapture(), please excute sudo nvpmodel -m 0 and sudo jetson_clocks. This fixes CPU cores at max clock and is the optimal solution for the use-case.