Issues for Image Matching in Structure from Motion

This paper discusses issues for matching point features in structure from motion, i.e., pose estimation and 3D reconstruction. As our building block are image triplets for which matches can be checked also geometrically, we discuss the issues for triplets based on partial or final results of a complete system for structure from motion. Particularly, we analyze and try to give answers to the following three questions: Shall all good matches be used or only the best matches? Does covariance information help? and finally – Is least-squares matching useful? We discuss the issues also from a theoretical point of view, but we mostly address them empirically by means of three suitable, i.e., difficult examples, as we found that answers to the above questions are strongly tied to the characteristics of real data.