Abstract
One of the key arguments against view-based methods as a model of navigation and spatial representation is that such models cannot (at present) explain the following performance: an observer (human, animal or robot) views two objects in one visual direction, one of which is closer to the observer than the other; the target objects disappear while other objects in the scene remain; the observer translates to a new location and then orients in turn towards the original targets (which remain invisible). Accurate performance on this task requires information (in some form, not necessarily metric) about the distance of the targets at the initial location. It is often assumed to require, in addition, a metric representation of the scene. Here we show through simulation that view-based information alone is adequate to carry out this task with reasonable accuracy over a range of conditions without the need to generate a metric reconstruction of the scene. We collect information about the change in relative visual directions (RVDs) of pairs of points in a 'flatland' 2D scene caused by a short-baseline translation (binocular stereo or motion parallax) collected at two locations (A and B) separated by a wide baseline. Using these RVDs and changes in RVDs, we show how it is possible to orient towards a subset of points that cannot be seen from a third location (C). In simple cases, such as C lying on the line AB, this is relatively straight-forward to achieve. We show, over a range of locations of C, the extent to which RVDs and changes in RVDs (optic flow and stereopsis) around A and B can support accurate pointing to unseen targets without the use of a metric map.
Meeting abstract presented at VSS 2012