Reconstruction of structured scenes from two uncalibrated images
Pattern Recognition Letters
| Hi-index | 0.00 |
Man-made environments possess many regularities which can beefficiently exploited for image based rendering as well as roboticvisual navigation and localization tasks. In this paper we presentan approach for automatic extraction of dominant rectangularstructures from a single view and show how they facilitate therecovery of camera pose, planar structure and matching acrosswidely separated views. Since in the presented approach therectangular hypothesis formation is based on a higher levelinformation encoded by the presence of orthogonal vanishingdirections, the dominant rectangular structures can be detected andmatched despite the presence of multiple repetitive structuresoften encountered in a variety of buildings. The different stagesof the approach are demonstrated onvarious examples of images ofindoors and outdoors structured environments.