Heuristic Search in 3D Object Recognition

摘要

One of the outstanding problems in visual recognition, natural language interpretation, and software automation, to name but a few, is that of discovering what to look for, or what is meant. The focus of this paper is ordering features for 3D object recognition using outdoor image sequences as inputs. Here, the problem is that of extracting relevant features. Furthermore, the problem is compounded by the need to rank the extracted features and combinations of features for relevancy. This paper takes a different view of the 3D object recognition problem. In addition to rotations, reflections, and simple image transformations, the paper presents the reader with a methodology for extracting relevant combinations of features. Our task is to apply heuristic search to find the most relevant sets of such features. The need for heuristics is implicit. The combinatorics is defined by the power set, which for just 20 features represents over one million combinations. Our solution will take advantage of the fact that the recognition problem is not NP-hard. Rather, it involves symmetry, which again differentiates it from an NP-hard problem. The exploitation of spatial and/or temporal symmetry will enable a far greater capability for the extraction of features than was previously possible. The sets of features can then be input to the 3D recognition algorithm, which for the observed instances will then be far better equipped to discriminate and classify similar instances.