Scene Reconstruction from Images

Abstract

The CAIP’99 invited lecture overviews the recent development in the scene reconstruction in the context of the Prague research contribution to it by the Center for Machine Perception.3D Scene Reconstruction from (uncalibrated) 2D images The second half of nineties witnessed a qualitative move from stereovision that remained for a long time in a one hundred old photogrammetric framework providing relation between two views only. The new impulse was the discovered trilinear and quadrilinear relation among views in projective geometry. Another impulse was the transition from Euclidean using calibrated cameras to projective reconstruction where uncalibrated cameras are sufficient. The observations of the scene provides extensive number (tenth of) images that yield qualitatively better results than before.The topic will be overviewed from the perspective of own achievements: (a) interpolation from two images [10], (b) relation among interpolation, extrapolation and reconstruction of a full 3D model [13], (c) projective reconstruction from three uncalibrated images [11], (d) introduction of the oriented projective reconstruction 14, 12, (e) selection of an optimal set of reference images [1], (f) search for correspondences if observer just translates or rotates [12], (g) the attempt to generalise a dense sequence correspondence to more general cases [9], (h) autocalibration from uncalibrated views 8, 4.Omni-directional Vision Epipolar geometry and egomotion estimation algorithm for central panoramic cameras was developed and presented 6, 3, 7. Design and image formation for newly defined central panoramic cameras have been studied [5]. New approach to mobile robot localization has been proposed [2]. The approach relies on a visual map comprising panoramic images which are represented in a rotationally invariant manner.KeywordsMotion EstimationPanoramic ImageEpipolar GeometryProjective ReconstructionMobile Robot LocalizationThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.