A new set of variable dimensional local shape descriptors for 3D registration is proposed and applied to 3D model building from range images. The descriptors are based on a large set of properties represented as high dimensional histograms. The novelty of the method is two fold: first, it offers a generalized platform for a large class of local shape descriptors; second, unlike previously devised de- scriptors that are of low dimensionality and compact size, these descriptors are high dimensional and highly discrim- inating. The new approach suggests investing more into de- scriptor generation and comparison and in return gaining a higher percentage of inliers in the set of hypothesized point matches across the images being registered. This in turn drastically reduces the required number of RANSAC iter- ations for finding the alignment between two images, as is confirmed by experimentation in a 3D model building appli- cation. It is also shown that the correct choice of properties can increase the effectiveness of feature correspondences, thereby increasing the possible acquisition angle between overlapping images.
展开▼
