Wide-baseline object interpolation using shape prior regularization of epipolar plane images

This paper considers the synthesis of intermediate views of an object captured by two calibrated and widely spaced cameras. Based only on those two very different views, our paper proposes to reconstruct the object Epipolar Plane Image Volume [1] (EPIV), which describes the object transformation when continuously moving the viewpoint of the synthetic view in-between the two reference cameras. This problem is clearly ill-posed since the occlusions and the foreshortening effect make the reference views significantly different when the cameras are far apart. Our main contribution consists in disambiguating this ill-posed problem by constraining the interpolated views to be consistent with an object shape prior. This prior is learnt based on images captured by the two reference views, and consists in a nonlinear shape manifold representing the plausible silhouettes of the object described by Elliptic Fourier Descriptors. Experiments on both synthetic and natural images show that the proposed method preserves the topological structure of objects during the intermediate view synthesis, while dealing effectively with the self-occluded regions and with the severe foreshortening effect associated to wide-baseline camera configurations.