The Team has not yet found a working solution to this problem that only seems to affect ZED X cameras.
We are working hard to fix it and a solution will be provided as soon as possible.
I changed the start method from multiprocessing in the main method to spawn:
mp.set_start_method('spawn', force=True)
Since ZED objects are not pickleable, I created two processes for the respective cameras that communicate via c_variables. With torch.multiprocessing you can also exchange images as tensors across processes.
Note: I work on two ZED2i cameras. So the problem regarding ZEDX is not solved.