Abstract:

PURPOSE: In a pair-of-pairs comparison experiment, Clarke et al. [2012] showed human observers two pairs of textures and required them to decide on which pair was more similar. Dong and Chantler [2013] further examined 51 sets of computational image features for estimating the humans’ perceptual judgements but they found that none performed well. A question hence arises: do humans estimate texture similarity in a consistent way? In this study, we conduct an experiment in order to answer it.
METHODS: The top 80 most inconsistent pairs of pairs of textures between humans’ judgements and the results obtained using features were selected. Only the pair that the majority of the 51 feature sets chose as dissimilar but the observers [Clarke et al., 2012] disagreed with this in each pair of pairs was used. An experiment was conducted using eight native English speakers. Given a pair of textures, the observer was required to describe the similarity of these using at least one of the 98 English words that Bhushan et al. [1997] used.
RESULTS: The occurrence frequencies of the 98 words were accumulated into a histogram. Since these words were divided into eleven clusters by Bhushan et al. [1997], the frequencies of the words in the same cluster were collapsed into a single bin. Considering the size of each cluster was di?erent, the collapsed frequency of each cluster was then normalized using its size. It was found that four clusters of words were more frequently used than the others, including “regular”, “lined”, “netlike” and “bumpy” (only one word was listed for each cluster).
CONCLUSIONS: Experimental results show that regularity, directionality and bumpiness mainly account for the perceptual judgements that human observers chose the 80 pairs in the pair-of-pairs comparison experiment [Clarke et al., 2012]. This finding suggests that humans estimate texture similarity in a consistent manner.

Key words: Texture similarity, texture properties, human perception