This report describes a preattentive recurrent neural network model which, given an image sequence as input, simultaneously achieves segmentation, occlusion-finding and optic flow mapping. Unlike previous models of these visual processes, this model doesn't employ a hierarchical, modular approach. Instead, the importance of heterarchically integrated processing is stressed, resulting in simultaneous output. In many conventional models, the extraction of image features (i.e. segmentation) and of optic flow are treated separately: image features are first extracted; and then these features are matched between frames. One of the novel aspects of this model is that it detects moving features solely as a consequence of determining optic flow. Another novelty is its detection of occlusion. Occlusion is a topic often ignored in optic flow analysis; however, it is argued here that occlusion detection is important for any visual motion system, for two main reasons. First, it supplies structural information (for later flow field analysis) on the relative depths of moving objects. Second, it provides information on the presence of stationary objects and surfaces, without the need for separate static image processing. These novel features of the model result in significant computational savings.
This paper is not available online