The Importance of Long-Range Interactions to Texture Similarity

Perceptually Motivated Image Features Using Contours

“…Dong et al [17], [19] introduced two evaluation methods for assessing the ability of computational features to estimate perceptual texture similarity: a pair-of-pairs comparison and a texture retrieval task. The two methods used two sets of human-derived data.…”

“…However, the task of perceptual similarity estimation is important and can be used in a number of applications, from measuring the perceived difference between the appearances of textures to simply ranking the results of search engines. Recently, Dong and Chantler [17] assessed 51 computational feature sets and found that the highest agreement rate with human data (obtained using free-grouping experiments [10]) was not greater than 61%, and that coincidently, none of the feature sets tested, exploited longer-range higher order statistics (i.e. HOS computed over areas >19×19 pixels).…”

“…However, higher order statistics are often computationally expensive to acquire, and are thus normally computed within limited spatial extent 1 . Hence "textons" and other vector quantisation methods are typically limited to 19×19 pixel neighbourhoods [17], [19].…”

“…We have found few texture feature sets that capture long-range (>19×19 pixels), aperiodic texture characteristics [17], [19]. It is well-known, however, that these characteristics have an important role in human visual perception [15], [25], [42], [44], [53], [55].…”

“…Second, we develop and test a novel feature set that exploits long-range HOS by encoding spatial and angular distributions of contour segments. We assess this feature set's performance against (1) the 51 feature sets that Dong et al [17], [19] examined and (2) four additional feature sets based on shape representation. We also test the feature set in the context of two other popular tasks: sketch-based image retrieval [21] and natural scene recognition [6].…”