Image Geometries
Deriving geometric representations from images of the humanly constructed world is a long-standing task in vision. The main objective is to characterize the surroundings that we are interacting with into rich geometric and semantic structures, such as wireframes and planar surfaces. Our goals is to detect these geometric primitives from single-view images using modern deep learning, without demanding large manually labeled datasets. We achieve this goal by incorporating geometric priors into learning such that built-in knowledge no longer needs to be learned from data.
CVL people: Yancong Lin
