Pose versus Correspondence Search

Most registration algorithms align data sets by finding approximately corresponding data features and then estimating the pose that aligns these. A problem that arises from this approach is the convergence to significantly misaligned local minima, which can happen when the data sets are initially far from correct alignment or slight mis-alignment when near to the global optimum. More recent research has started search in the pose space instead of the correspondence space and seem to be finding a broader range of initial poses that still lead to convergence near the correct alignment (11,55,25,9). So here essentially we have to make clear how to measure the quality of correspondences independent of the algorithms used to estimate the pose. At the beginning of the registration most correspondences are not correct, and the converse problem arises as how to measure the quality of the pose from such inaccurate correspondences. In fact, we may never know whether estimated correspondences are exactly right or not and this has an obvious effect on the quality of the estimated pose parameters.