Towards Aspect Invariant Feature Sets for Characterizing Three Dimensional Objects

摘要

The problem of characterizing three dimensinal objects with a minimum number of feature sets is addressed. Previous research has been plagued by features which are a function of aspect angle and so efforts have concentrated on characterizing an object with several hundred prototype feature sets. In this work it is demonstrated that in many cases, the silhouette obtained from one view of the object can be derived from a linear transformation of the silhouette from another view. As a result of this relationship, a single set of moments which is invariant to such a general linear transformation can be used to characterize many views of the same object and hence the number of prototype feature sets required to specify an object is reduced. In addition, it is demonstrated that for some objects it is advantageous to partition the object into regions in order to find the region of the object which is least dependent upon aspect angle. (Author)