Global Transformation Estimation via Local Region Consensus

This work presents a novel method for accurate estimation of perspective transformations in images. Recent publications have shown that considering a rich geometric transformation model of salient local regions, is extremely bene cial for the purpose of point matching.Yet, those methods are not used extensively for two reasons. First, because they are computationally more demanding, second reason is that the estimation of the geometric transformations are of limited accuracy and therefore the usage of them is limited. Moreover, in typical scenes, signi cant portions of the scene are low-textured and cannot be considered salient. Thus their transformation between images can be recovered only by related global transformations. It has been shown that considering a few local point matches in consensus can be very useful in estimating global transformation and rejecting outliers. Though these methods do not typically utilize information from local region transformations. In this work, we propose a mechanism that forges a consensus between local region correspondences and their underlying geometry , that allows the estimation of richer global transformations. We demonstrate how this idea enables the estimation of perspective transformations of planes, and locating many new point matches on practically texture-less areas of the scene.