It seems the camera could use optical flow to get a baseline reading and calibrate the inertial frame offsets. They don't need to point accurately for a long time?

Or maybe that is the method they assume for the second solution and they calculated that it's infeasible.