Recognition strategy generation for pose estimation of multiple 3-dimensional objects

An interpretation tree can be regarded as a recognition strategy precompiled so that it can offer fast run-time performance. Previous interpretation trees are limited to recognizing only one type of object. The authors present a method of generating interpretation trees capable of recognizing objects of various types. The problem which may arise is the large number of aspects needed to put all the objects in one tree. It might cause the compilation time for generating an interpretation tree to be prohibitively long. Two approaches are suggested to avoid this problem. One is a concept of super-aspect which can be regarded as an hierarchical representation of aspects. This representation makes it possible to reduce the compilation time by handling a smaller number of aspects at a time. The other is the reduction of the number of aspects itself by utilizing the rotational and mirror symmetry objects might have. Interpretation trees constructed in this manner are presented for various objects and tested through experiments with light-stripe range images