Belief-Propagation on Edge Images for Stereo Analysis of Image Sequences

Abstract

The history of stereo analysis of images dates back more than one hundred years, but stereo analysis of image sequences is a fairly recent subject. Sequences allow time-propagation of results, but also come with particular characteristics such as being of lower resolution, or with less contrast. This article discusses the application of belief propagation (BP), which is widely used for solving various low-level vision problems, for the stereo analysis of night-vision stereo sequences. For this application it appears that BP often fails on the original frames for objects with blurry borders (trees, clouds, . . . ). In this paper, we show that BP leads to more accurate stereo correspondence results if it is applied on edge images, where we have decided for the Sobel edge operator, due to its time efficiency. We present the applied algorithm and illustrate results (without, or with prior edge processing) on seven, geometrically rectified night-vision stereo sequences (provided by Daimler AG, Germany).