Object recognition from range data prior to segmentation

摘要

An unconventional approach to range image understanding is proposed, unconventional in the sense that object matching and recognition is performed before the image is segmented. The motivation for this is our own perception of problems, which we analyse in the paper, arising from the conventional paradigm of recognition after segmentation. The proposal consists of two levels: the representation and the computational levels. We illustrate how to represent scenes and models using attributed relational graphs (ARG). The ARG representation describes surface points, rather than segmented primitives, as its nodes, their invariant properties and invariant relations between them. Object matching and recognition, as well as image segmentation, is determined by the optimal point-wise mapping from the image ARG and the model ARG. We formulate the problem of finding such mapping as one of constrained optimization. The solution is computed using the relaxation labelling method, and some experiments of such optimal mapping are presented.