Improved Inbetweening

The objective of this work is to unify an image-centric 2-D view of the world and a visioncentric 3-D view of the world in the context of inbetweening/view interpolation. This is achieved by incorporating epipolar geometry into an existing inbetweening algorithm [8], which is built on a Bayesian framework. Issues include estimation of the fundamental matrix, specification of the new view and how to incorporate the geometry. The resulting hybrid algorithm exploits the geometry by inserting an additional prior, an epipolar constraint prior, into the existing Bayesian framework. When this hybrid algorithm is used to produce new views, the improvement is manifested in better treatment of occlusion. It is expected that, through improving the accuracy of the new view specification and through more experiments in exploitation of the geometry, further improvements could be achieved.