Editor's Choice Article: Image-consistent patches from unstructured points with J-linkage
Image and Vision Computing
Hi-index | 0.00 |
Dense 3D reconstruction in man-made environments has to contend with weak and ambiguous observations due to texture-less surfaces which are predominant in such environments. This challenging task calls for strong, domain-specific priors. These are usually modeled via regularization or smoothness assumptions. Generic smoothness priors, e.g. total variation are often not sufficient to produce convincing results. Consequently, we propose a more powerful prior directly modeling the expected local surface-structure, without the need to utilize expensive methods such as higher-order MRFs. Our approach is inspired by patch-based representations used in image processing. In contrast to the over-complete dictionaries used e.g. for sparse representations our patch dictionary is much smaller. The proposed energy can be optimized by utilizing an efficient first-order primal dual algorithm. Our formulation is in particular very natural to model priors on the 3D structure of man-made environments. We demonstrate the applicability of our prior on synthetic data and on real data, where we recover dense, piece-wise planar 3Dmodels using stereo and fusion of multiple depth images.