Feature point correspondences: A matching constraints survey.

Abstract

The correspondence problem remains of central interest in image analysis. Matching images is a fundamental step in computer vision applications like the recovery of 3D scene structures, the detection of moving objects, the synthesis of new camera views and many others. Several algorithms are proposed in the literature to solve the difficult problem of feature correspondences between image pairs. They use different approaches and constraints to improve the quality of the resulting matching set. Although several solutions have been proposed in the past, finding good point correspondences between stereo pairs or in image sequences is still a very challenging task. In this work, we present a survey of different types of matching constraints and propose an empirical evaluation of their performance. Experiments on several pairs of images were conducted where we show in graph forms the behavior of each one of the constraints in terms of elimination of false matches. The validation process of a good match is based on the epipolar geometry associated with the image pairs. This geometrical information is obtained by calibrating the image pairs using an existing calibration software.