Matching and Pose Refinement with Camera Pose Estimates

Abstract

This paper describes novel algorithms that use absolute camera pose information to identify correspondence among point features in hundreds or thousands of images. Our incidence counting algorithm is a geometric approach to matching; it matches features by extruding them into an absolute 3-D coordinate system, then searching 3-D space for regions into which many features project. The absolute pose estimates reported by our instrumentation are accurate, but not perfect. Thus, we also consider the problem of refining these pose estimates, given feature matches from a set of images. We describe a pose refinement algorithm which decouples translation (position) estimates from rotation (attitude) estimates, and can incorporate matches from many hundreds or thousands of images. 1 Introduction Many 3-D reconstruction algorithms rely on a matching or correspondence step to identify constraints corresponding to the scene geometry; these constraints are used to guide the 3D reconstruction proc...