Your camera would only run in one resolution/format/fps at once.
So you would try to capture with highest required resolution, and create a second frame by resizing the high res one. You would therefore have 2 images, one for each net.
AFAIK, there is no python binding for cudaResizeRGBA(). For your case, C++ API may be easier.
You would allocate a ZeroCopy array for your second image in init, and in the loop after capture you would create the smaller image with cudaResizeRGBA().
For python, there is info in this topic that may help, but may not be as efficient as CUDA.
I have poor experience with python, but maybe someone else would be able to tell, if possible, how to add python binding for cudaResizeRGBA().