Proceedings Article | 18 February 2010
KEYWORDS: Visualization, Image compression, Visual process modeling, Statistical modeling, Visual system, Spatial frequencies, Data modeling, Image processing, Systems modeling, Databases
Factors governing human preference for artwork have long been studied but there remain many holes in our
understanding. Bearing in mind contextual factors (both the conditions under which the art is viewed, and the state of
knowledge viewers have regarding art) that play some role in preference, we assess in this paper three questions. First,
what is the relationship between perceived similarity and preference for different types of art? Second, are we naturally
drawn to certain qualities-and perhaps to certain image statistics-in art? And third, do social and economic forces
tend to select preferred stimuli, or are these forces governed by non-aesthetic factors such as age, rarity, or artist
notoriety? To address the first question, we tested the notion that perceived similarity predicts preference for three
classes of paintings: landscape, portrait/still-life, and abstract works. We find that preference is significantly correlated
with (a) the first principal component of similarity in abstract works; and (b) the second principal component for
landscapes. However, portrait/still-life images did not show a significant correlation between similarity and preference,
perhaps due to effects related to face perception. The preference data were then compared to a wide variety of image
statistics relevant to early visual system coding. For landscapes and abstract works, nonlinear spatial and intensity
statistics relevant to visual processing explained surprisingly large portions of the variance of preference. For abstract
works, a quarter of the variance of preference rankings could be explained by a statistic gauging pixel sparseness. For
landscape paintings, spatial frequency amplitude spectrum statistics explained one fifth of the variance of preference
data. Consistent with results for similarity, image statistics for portrait/still-life works did not correlate significantly with
preference. Finally, we addressed the role of value. If there are shared "rules" of preference, one might expect "free
markets" to value art in proportion to its aesthetic appeal, at least to some extent. To assess the role of value, a further
test of preference was performed on a separate set of paintings recently sold at auction. Results showed that the selling
price of these works showed no correlation with preference, while basic statistics were significantly correlated with
preference. We conclude that selling price, which could be seen as a proxy for a painting's "value," is not predictive of
preference, while shared preferences may to some extent be predictable based on image statistics. We also suggest that
contextual and semantic factors play an important role in preference given that image content appears to lead to greater
divergence between similarity and preference ratings for representational works, and especially for artwork that
prominently depicts faces. The present paper paves the way for a more complete understanding of the relationship
between shared human preferences and image statistical regularities, and it outlines the basic geometry of perceptual
spaces for artwork.
