Feature Points Densification and Refinement

摘要

A large part of computer vision algorithms and tools rely on feature points as an input data for the future computations. Given multiple views of the same scene, the features, extracted from each of the views can be matched, establishing correspondences between pairs of points and allowing their use in higher-level computer vision applications, such as 3D scene reconstruction, camera pose estimation and many others. Nevertheless, two matching features often do not represent the same physical 3D point in the scene, which may have a negative impact on the accuracy of all the further processing. In this work we suggest a feature refinement technique based on a Harris corner detector, which replaces a set of initially detected feature points with a more accurate and dense set of matching features.