An Optimal Object Representation For Aspect Classification
See P. Toh, Mark Henderson · 2005
Ab s t r a c t - - Aspect c lassification is the t ask of approximating the precise pose of an unknown object by classifying the unknown as one of the aspects in the aspect graph. The precise pose of the unknown is then determined by using the aspect as a starting point and solving for linear change. Aspect classification is equivalent to the task of solving for aspect change. This task r equires intense searching of aspect graph and matching of sets of image features with the sets of model features. The process can be speedup by organizing and indexing the i nformation available in t he aspect graph in a lookup table. A data structure is designed based on which an indexing mechanism is realized. We introduce the concept of discriminators and aspect d iscrimination t rees (ADT) which makes a given set of aspects completely unambiguous. In general, more than one ADT exists for a given set of aspects and associating set of discriminators. We present an A*-based algorithm OED for constructing an optimal and complete ADT. A path f rom the root node to a leaf node in an ADT indicates a successful classification of the unknown. The constructed ADT with t he redundant and repetitive information eliminated is stored and retrieved at classification time. Finally, we analyze the s torage r equired to store a constructed ADT.