中国科学院自动化研究所机构知识库

A new and effective dense 3D scene reconstruction algorithm is proposed in this work. At first, in order to speedup the reconstruction process and reduce possible unnecessary interference by unrelated regions, such as sky, ground, etc., to the desired architectural scene, a semantic image segmentation step is carried out at first to mask those uninterested parts. After that, an initial depth map is obtained by Daisy-feature based 3D

point propagation. To tackle the difficult dense reconstruction problem of poorly textured regions and slanted surfaces, the problem is formulated as a plane inference problem under the MRF energy minimization framework to assign a possible planar patch to each image superpixel. More specifically, at first, the plane principal directions and possible non-principal are clustered by the affinity propagation from the reconstructed point clouds, then the plane inference is carried out along the principal directions first, followed by the non-principal directions. In each of the above two plane inference steps, a Graph Cuts based algorithm is employed to minimize a properly designed cost function which takes into account a variety of factors, such as appearance similarity, the constraints from the

initial depth map, spatial planes priors, as well as geometric relations between spatial planes etc. Experiments on the standard data set as well as our own data set show that the our proposed algorithm can satisfactorily handle many challenging factors, e.g. illumination variation, perspective distortion, poorly textured regions, etc.,

to efficiently recover complete dense depth map of buildings.