OPTIMAL PHOTO HULL RECOVERY FOR THE IMAGE-BASED MODELING

摘要

In the voxel-based scene reconstruction, each voxel projects to an image as a polygonal region. Therefore, the estimated color of each voxel is differently compressed and distorted according to the scene textureness. In addition, the sensor noise and modeling errors increase the uncertainty in the occupancy decision of a voxel. All of these are hard to be considered in the voxel-centered statistical modeling that is commonly adopted in previous approaches. In addition, this kind of modeling is inadequate for the optimization-based approach. As a result, even though the recovered scene successfully regenerates the input images, the true scene geometry is not always preserved in the recovered scene. To recover the optimal photo hull that satisfies the theoretical photo hull theorem and additionally preserves the true scene geometry, this paper introduces a novel pixel-centered statistical modeling method and a simple optimization-based scene reconstruction paradigm. To model both characteristics of voxel and various uncertainties, this paper utilizes the relationship between a pixel and its occupying candidate voxels. And any kind of ad hoc factor is unnecessary in the proposed paradigm.