A novel algorithm for refinement of vision-based two-view pose estimates

A novel method is developed to obtain a refined estimate of relative position and orientation (pose) from two views captured by a calibrated monocular camera. Due to the typically large number of matched pairs of feature points available, many estimates of the pose are possible by taking minimal sets of feature points. Among such estimates, the proposed method selects a subset that has low estimation error, and averages them in an appropriate manner to provide a refined estimate. Preliminary experiments on the image data show that the proposed method provides a more accurate pose estimate than refinement using a least squares method.