Elements of Style: Learning Perceptual Shape Style Similarity

摘要

The human perception of stylistic similarity transcends structurernand function: for instance, a bed and a dresser may share a commonrnstyle. An algorithmically computed style similarity measure thatrnmimics human perception can benefit a range of computer graphicsrnapplications. Previous work in style analysis focused on shapesrnwithin the same class, and leveraged structural similarity betweenrnthese shapes to facilitate analysis. In contrast, we introduce thernfirst structure-transcending style similarity measure and validaternit to be well aligned with human perception of stylistic similarity.rnOur measure is inspired by observations about style similarity in artrnhistory literature, which point to the presence of similarly shaped,rnsalient, geometric elements as one of the key indicators of stylisticrnsimilarity. We translate these observations into an algorithmicrnmeasure by first quantifying the geometric properties that make humansrnperceive geometric elements as similarly shaped and salientrnin the context of style, then employing this quantification to detectrnpairs of matching style related elements on the analyzed models,rnand finally collating the element-level geometric similarity measurementsrninto an object-level style measure consistent with humanrnperception. To achieve this consistency we employ crowdsourcingrnto quantify the different components of our measure; we learn thernrelative perceptual importance of a range of elementary shape distancesrnand other parameters used in our measurement from 50K responsesrnto cross-structure style similarity queries provided by overrn2500 participants.We train and validate our method on this dataset,rnshowing it to successfully predict relative style similarity with nearrn90% accuracy based on 10-fold cross-validation.