Utilizing temporal associations for view-based 3-D object recognition

The authors propose an architecture for the recognition of three-dimensional objects on the basis of viewer-centered representations and temporal associations. Motivated by biological findings and by successful computational implementations they have chosen a viewer-centered representation scheme. In contrast to other implementations, special attention is paid to the temporal order of the views, which proves useful for learning and recognition purposes. Their recognition system combines different kinds of artificial neural networks into a four stage architecture: preprocessing by a Gaborjet transform is followed by an extended dynamic link matching algorithm which implements recognition and learning of the view classes. A STORE network records the temporal order of the views by transforming a sequence of view classes into an item-and-order coding. Subsequently, a Gaussian-ARTMAP architecture is used for the classification of the sequences and for their mapping onto object classes by means of supervised learning. The presented results demonstrate that the system is capable to autonomously learn and to discriminate similar objects. Additionally, the examples show how the utilization of the temporal context improves object recognition by making ambiguous views manageable and facilitating an increased insensitiveness against misclassifications.