CARVING 3D MODELS FROM UNCALIBRATED VIEWS
Miguel Ángel Sainz, Nader Bagherzadeh, Antonio Susín · 2002
In this paper we present an automatic method for the reconstruction of a 3D volumetric representation of real world scenes from a set of multiple uncalibrated images. The process is divided in two steps (1) an automatic calibration of the cameras and (2) a scene reconstruction consistent with the input views. The calibration of the cameras is performed using automatically tracked 2D features, and consists in calculating a projective approximation and upgrading it to an Euclidean structure by computing the projective distortion matrix in a way that is analogous to estimate the absolute quadric. Moreover, in contrast to other approaches our process is essentially a linear one. The underlying technique is based on the Singular Value Decomposition (SVD) and the process is enhanced with a careful study of the rank of the matrices involved in order to get the excellent results shown in the paper. The volumetric reconstruction of the scene is performed using an improved voxel carving algorithm. The result is a voxel-based model of the external surface of the physical objects present in the scene. Optimized data structures and graphics hardware acceleration are used to achieve a substantial reduction in computation time. Furthermore, the spatial information obtained from the the camera calibration process about the 2D tracked measurements is used to automatically set the internal thresholds of the carving algorithm, achieving a full automation of the method.