Abstract
Style is a multidimensional attribute of an artwork. While art historians may categorize artworks based on high-level concepts, such as the historical period, this level of description may not align with low-level visual properties of style that are extracted by the visual system. Here, we attempt to reveal the relevant dimensions of style by asking which visual features determine how participants categorize artworks. Using a natural grouping task, participants were asked to hierarchically sort artworks from eight artistic periods into eight groups based on the criteria of their choice. They were then asked to provide labels for each sub-group they created. Results of a thematic analysis revealed that participants primarily sorted on high-level concepts such as semantic image content (e.g. people) and broad style themes (e.g. modern) while lower level-visual features such as colour and texture were used only at lower hierarchical levels. In addition, we conducted multidimensional scaling analysis (MDS) to quantitatively determine which features participants used to categorize the paintings. The results mirrored that of the thematic analysis with the most relevant dimension separating the paintings by broad high-level elements such as historical period (old-new). In an attempt to maximize the contribution of lower-level stylistic elements, we repeated the same experiment preventing participants from relying on semantic content as a categorical dimension. As expected, results of the thematic analysis showed participants relying more heavily on colour and image structure at the first level of grouping, although the MDS showed no clear single dimension emerging from the analysis. These findings are in line with previous work (Augustin & Leder, 2006; Wallraven et al., 2009) emphasizing the high-level processing that the average observer makes when viewing artworks, forming meaningful and consistent categories. However, these consistencies are largely based on image content rather than low-level visual characteristics.
Meeting abstract presented at VSS 2016