A similarity measure for illustration style

Abstract

Figure 1: The leftmost composition is generated by selecting from a dataset of 200K clip art searched with keywords: dog, tree, sun, cloud, and flower. Unfortunately, the styles of the clip art are inconsistent. Our style similarity function can be used to re-order the results of a search by style. The next three scenes are generated by fixing one element, and then searching for stylistically similar clip art with the above keywords. In each case, the additional clip art were chosen from the top twelve returned results (out of thousands). AbstractThis paper presents a method for measuring the similarity in style between two pieces of vector art, independent of content. Similarity is measured by the differences between four types of features: color, shading, texture, and stroke. Feature weightings are learned from crowdsourced experiments. This perceptual similarity enables style-based search. Using our style-based search feature, we demonstrate an application that allows users to create stylisticallycoherent clip art mash-ups.

Style compatibility for 3D furniture models

“…Nonetheless, our results are comparable to those obtained in other subjective evaluations for aesthetic suggestion interfaces [O'Donovan et al 2014;Garces et al 2014]. …”

“…Previous work has used the L1-norm to sparsify weight vectors [Garces et al 2014]. However, applying the L1-norm to the entries of the embedding matrix (R(Wc) = |Wc|1), would only have the effect of sparsifying individual matrix entries, rather than eliminating entire dimensions.…”

“…Our work is related to similarity metric learning in other domains. Researchers have used crowdsourced data to learn similarity metric for fonts [O'Donovan et al 2014] and illustration styles [Garces et al 2014]. Relative attributes have been learned from visual features for image analysis [Parikh and Grauman 2011] and from shapes for 3D model creation .…”

“…Previous work [Kulis 2012;O'Donovan et al 2014;Garces et al 2014] has employed distance functions of the form…”

“…Given a distance function, learning proceeds similar to previous work [O'Donovan et al 2014;Garces et al 2014]. The probability that a rater evaluates object A as more compatible to B than to C is modeled as a logistic function:…”