A Unified Approach to Image Matching and Segmentation in Stereo, Motion, and Object Recognition via Recovery of Epipolar Geometry

Gang Xu
Computer Vision Laboratory
Department of Computer Science
Ritsumeikan University



Abstract

In this paper I try to show that through recovering epipolar geometry we can provide a unified approach to the problems of image matching and segmentation in stereo, motion, and object recognition, which have been treated separately so far. Stereo matching has been known as a 1D search problem, while matching in motion and object recognition have been known as 2D search problems. I show that by recovering epipolar geometry underlying the images, the correspondence search problems in motion and object recognition can also be changed to be 1-dimensional. I propose a new approach to recovering epipolar geometry for multiple rigid motions using feature points in two uncalibrated images. Using the recovered epipolar equations, the edge images are then easily matched and segmented. Examples are shown for matching and segmenting motion images with multiple rigid motions, and for matching model view against input view and localizing the 3D object in the input view.

Keywords: clustering, correspondence, epipolar geometry, motion, object recognition, robust estimation, stereo