Inverse validation for accurate range image registration with structured data

Automatic range image registration is a fundamental yet extremely difficult problem in machine vision. In this paper, we extend a promising registration algorithm, the geometric iterative closest point (GICP), for structured image data. The extended algorithm is based on a relaxation method combining the motion estimation results from the possible correspondences established by the GICP algorithm and those that are further evaluated by an inverse validation procedure. A comparative study based on real images has shown that the extended algorithm is more accurate and robust than the original algorithm.