Articulated Binocular Vision

Abstract

The previous two chapters have shown that, in order to accomplish figure/ground segmentation based on the relative motion between objects in a scene, four distinct images are required to provide sufficient constraints on the four ‘unknowns’ of foreground, background, and their respective velocities or displacements. This is true irrespective of whether the relative object motion arises from independent movement of the objects or from camera motion. Essentially, a quadocular configuration is required, i.e. we need four images acquired at four equally-spaced positions x₀, x₁, x₂, x₃. If we are dealing with a translating camera rather than four distinct fixed cameras, then x _i = vt _i where v is the camera velocity and t _i represents the time at which image i is acquired. Often, a binocular configuration is desirable. The question which arises then is: Can we effect image segmentation if we have two samples from each of two cameras? That is, can we combine spatial and temporal sampling in a unified spatio-temporal framework? Intuitively, it would seem that we can and the purpose of this chapter is to show formally how this can be achieved.