Constructing a Multivalued Representation for View Synthesis

摘要: A fundamental problem in computer vision and graphics is that of arbitrary view synthesis for static 3-D scenes, whereby a user-specified viewpoint the given scene may be created directly from representation. We propose novel compact representation this purpose called multivalued (MVR). Starting with an image sequence captured by moving camera undergoing either unknown planar translation or orbital motion, MVR derived each preselected reference frame, then used to synthesize views scene. The itself comprised multiple depth intensity levels which k-th level consists points occluded exactly k surfaces. To build respect particular dense maps are first computed all neighboring frames frame. combined together into single map, where organized occlusions rather than coherent affine motions. This grouping facilitates automatic process determine number helps reduce artifacts caused An iterative multiframe algorithm presented estimation both handles low-contrast regions produces piecewise smooth maps. Reconstructed as well flyarounds real scenes demonstrate effectiveness approach.