Euclidean Reconstruction from Uncalibrated Views

The possibility of calibrating a camera based on the identification of matching points in several views of a scene taken by the same camera has been shown by Maybank and Faugeras ([13, 4]). Using techniques of Projective Geometry they showed that each pair of views of the scene can be used to provide two quadratic equations in the five unknown parameters of the camera. A method of solving these equations to obtain the camera calibration has been reported in [13, 4, 12] based on directly solving these quadratic equations using continuation. It has been reported however that this method requires extreme accuracy of computation, and seems not to be suitable for routine use. In addition with large numbers of cameras (more than three or four) this method threatens to be unworkable, because the number of potential solutions grows exponentially in the number of views.