MDL estimation of a dense map of relative depth and 3D motion from a temporal sequence of images

Abstract

The goal of this study is to evaluate a dense, boundary preserving estimate of relative depth and motion in space from a temporal sequence of monocular images. The scheme we propose is based on the minimisation of a global energy function and the Minimum Description Length (MDL) principle: it seeks a dense estimate of greatest conformity to the changing image and to a structural model. Conformity is measured in terms of the number of bits to code its description. This MDL formulation refers explicitly to boundaries of depth and motion without introducing variables to name their positions, and underlying computations of optimisation reduce to Jacobi-like iterations. The formulation is direct, insomuch as relative depth and motion in space are estimated without prior estimation of optical flow. We derive it from the expression of an MDL estimate of optical flow by writing optical velocity in terms of depth and motion in space, assuming environmental objects are rigid. Viewers strongly perceived depth from stereoscopic images constructed from the scheme’s output.