Can spatio-temporal energy models of motion predict feature motion?

Current "spatio-temporal energy" models of how we perceive pattern motion have been very successful in helping us to understand the mechanisms of motion perception. Although they have been supported by a large number of physiological and psychological studies, they have so far not provided a complete explanation for a number of results. These results emerge from experiments concerned with predicting perceived motion direction from patterns comprising two or more components. It has been suggested that these results are more consistent with an earlier type of model based on the motion of two-dimensional features. This paper briefly describes how three generic spatio-temporal energy models have been extended to predict motion derived from two-component stimuli. A new model is then presented that utilises similar architecture to the two-stage spatial-temporal energy model proposed by Adelson and Movshon (Nature 300 (1982) 523). The first stage is a spatial temporal filtering stage and the second stage computes the intersection of constraints (IOC), an important constraint used in combining motion information across two or more components. In the model presented here the second stage is different. A directional spatial second derivative is used to extract zero-crossings at the component level, i.e. gratings. If any zero-crossing falls in the same spatial position for two or more components its displacement is tracked using a nearest neighbour match. Tracking these 'intersecting zero-crossings' essentially computes the IOC but also provides other properties that predict non-IOC motion, and second-order component motion. Surprising new insights are described into how current spatio-temporal energy models may also account for these results. However, unlike the model presented here, they rely on operations carried out on the two-dimensional pattern.