More realistic input images are processed by the VIEWNET architecture (View Information Encoded with NETworks) described by Grossberg and Bradski (1995). The view based model includes a biologically motivated preprocessing chain to convert the input images into a representation invariant under illumination changes, translation, plane rotation, and scaling. The classification of the resulting patterns is done by a Fuzzy-ARTMAP network (Carpenter et al. (1992)), an artificial neural network stemming from the adaptive resonance theory (ART). While this study hints at the advantages resulting from the consideration of view sequences instead of single images, it still neglects the order in which the views appear.
Such an evaluation of the serial information can be found in Seibert and Waxman (1992). Their system is able to create transition matrices from view sequences and thus offers a method for the automated construction of aspect graphs as defined by Koenderink and van Doorn (1979). However, the edges of an aspect graph indicate the transition between merely two views. The assembly of longer sequences requires the implicit assumption of transitivity along edges. Although the connections are augmented by the relative frequency of the according transition, detailed information about longer sequences is not provided because of the missing serial relations.
Darrell and Pentland (1993) present another system for the processing of view sequences. The employed image processing algorithms, however, are quite simple. Moreover, training and recognition require an alignment, called dynamic time warping, of the current input sequence and all the known sequences.