Global motion compensation and video mosaicing using different 2-D motion models

This paper gives a comparative study of the suitability of parametric 2-D motion models for global motion description. These investigations cover different aspects of short-term description, as used for global motion compensation, and long-term description, as used for video mosaicing. The presented results show, that linear models are mostly sufficient for a short-term description of global motion. Only in case of extreme camera operation (pan) and a complex depth structure of the background, application of non-linear models leads to significantly better results. For a long-term description of global motion and video mosaicing, the non-linear models clearly out-perform linear models in terms of description accuracy. On the other hand, linear models are more robust and stable in case of extremely large global motion.