Perception, 27(suppl.): 205-206. [ECVP 1998; Oxford, England.]
We investigated the representation of textured patterns underlying the detection of scene changes. Stimuli consisted of a number of texture patches, each randomly selected from two possible textures. Display of the patches was interrupted by display of a blank field at fixed intervals. In the Change condition, one of the patches would "flip" from one texture to the other during each blank interval. Subjects indicated whether or not a change occurred, and we measured response time as a function of the number of patches in the display. We tested a number of possible pairs of textures, and also ran several control experiments in which we instead changed an attribute of a simple "object," such as a rectangular bar.
We found that even large changes in the 2nd order statistics of a texture may result in slow search times, on the order of 200 ms/item. This includes large rotations of the texture, and scale changes of a factor of 2 or more. Similar changes in simple "objects" are detectable in 100 ms/item or less. However, some texture changes lead to more efficient search, on the order of 100 ms/item or less, such as the change from regular to random patterns, or from single to multiple orientations.
These results suggest a very different representation of texture than the results of texture-segmentation experiments. The determining factor for whether a change will be easily detected remains elusive. Perhaps a change in texture classification may be easily detected, e.g. a change from oriented texture to non-oriented, or from regular to random; or there may be a lower-level explanation. Performance seems to be worse for transformations of texture than of objects, though there's some evidence that the relevant dimension is the degree of complexity of the pattern, rather than whether the patterns are "objects" or "textures."