Accurate Object Reconstruction by Statistical Moments

Statistical moments can offer a powerful means for object description in object sequences. Moments used in this way provide a description of the changing shape of the object with time. Using these descriptions to predict temporal views of the object requires efficient and accurate reconstruction of the object from a limited set of moments, but accurate reconstruction from moments has as yet received only limited attention. We show how we can improve accuracy not only by consideration of formulation, but also by a new adaptive thresholding technique that removes one parameter needed in reconstruction. Both approaches are equally applicable for Legendre and other orthogonal moments to improve accuracy in reconstruction.