Vision based navigation systems have not been widely deployed, despite
the great deal of effort that has gone into researching the structure
from motion and vision based control. One reason for this is the
prevalence forced-choice parameter estimation, which use tools such as
maximum likelihood estimation, to estimate particular parameters such
as the structure of the scene or the motion of the camera. The
alternative to this is to maintain a probability distribution over the
parameter solution set. This project explores uses for such
representations of scene structure and camera motions, especially
questions in rendering models of objects with uncertain structure.
Approach
The approach is to replace the depth map as the desired output of
stereo or structure from motion.