Abstract
Visual attention lets us track objects based solely on their gradually changing appearance, without the help of distinct spatial location. Blaser and colleagues' ('00) memorably demonstrated this ‘tracking through feature space’ by superimposing two patterns with rapidly altering orientation, hue, and spatial frequency. Recurrent interactions in the relevant feature maps provide a plausible mechanism for maintaining object identity during rapid feature drift. Together with attentional bias, a balance between cooperative and competitive interactions may ensure the continued saliency of a rapidly changing pattern. To test this hypothesis, we extend the paradigm of Blaser and colleagues to coherent pattern motion (Adelson, Movshon, '82). Our patterns are composed of discrete log-Gabor wavelets and are characterized by pattern direction ?, spatial frequency ?r, temporal frequency ?t, and range of wavelet orientations ?. To ensure coherence, the speed vi of an individual wavelet varies with its orientation ?i (vi / vmax = cos(?i-?) ? cos ?). Contrast is fixed at 5 ∞ detection threshold for each ?r and ?t. In 5 s trials, two transparently superimposed patterns change rapidly in all 4 properties, with one ‘closest approach’ in feature space near the mid-point. Observers report which of the two final patterns is ‘descended from’ that cued initially. Tracking performance is limited by the patterns' distance at closest approach, as well as by the rates of feature change. Surprisingly, tracking improves with increasing range of wavelet orientations ? (i.e., a pattern of randomly oriented wavelets is easier to track than one of parallel wavelets). Comparing the quantitative results to those obtained with flashed wavelet arrays (300ms presentation, see Pastukhov, Festman, Braun, VSS '04) reveals the effect of pattern persistence on the perception of coherent motion. Blaser, Pylyshyn, Holcombe (2000) Tracking an object through feature space. Nature 408: 196-199.
EU IST-2001-38099