GENERATING 3D SCENE ACCORDING TO IMAGE TAKEN AT DIFFERENT ANGLES WHEN INTRINSIC PARAMETERS OF CAMERAS THAT TOOK THE IMAGES ARE UNKNOWN

摘要

A method and system for the accurate and efficient estimation of the fundamental matrix (FM) are disclosed. The estimation is carried out based on a number of noisy correspondences. It works by incorporating linear approximations of the unity and singularity constraints while minimizing either the algebraic cost function (as in the traditional Eight-Point (8P) scheme) or the weighted algebraic cost function (as in the traditional Weighted Eight-Point (W8P) scheme). When the algebraic cost function is adopted, the estimation gives results that are more accurate than those obtained by the 8P scheme while running almost as fast. When the weighted algebraic cost function is adopted, the estimation not only gives results that are more accurate than those obtained by the W8P scheme but also runs twice faster. The preferred embodiments make use of the linear Taylor series approximations of the unity and singularity constraints. The linear approximations are taken at the current best estimate of the FM. The resulting linear system of equations is then solved for an improved FM estimate. This process is repeated until the iterative procedure converges.