In this work, we address the problem of refining the geometry of local image
features from multiple views without known scene or camera geometry. Current
approaches to local feature detection are inherently limited in their keypoint
localization accuracy because they only operate on a