How do visual and conceptual factors predict the composition of typical scene drawings?

Abstract

Imagine you are asked to draw a typical bedroom, what would you put on paper? Your choice of objects is likely to depend on visual occurrence statistics (i.e., the objects present in previously encountered bedrooms) and semantic relations between objects and scenes (i.e., the semantic relationship between the bedroom and its constituent objects). To investigate how these two factors contribute to the composition of typical scene drawings, we analyzed 1,192 drawings of six indoor scene categories, obtained from 303 participants. For each object featured in the drawings, we estimated its visual occurrence frequency from the ADE20K dataset of annotated scene images, and its semantic relatedness to the scene concept from a word2vec language processing model. Across all scenes of a given category, generalized linear models revealed that visual and conceptual factors both predicted the likelihood of an object featuring in the scene drawings, with a combined model outperforming both single-factor models. We further computed the visual and semantic specificity of objects for a given scene, that is, how diagnostic an object is for the scene. Object specificity offered only weak predictive power when predicting the selection of objects, yet even infrequently drawn objects remained diagnostic of their scenes. Taken together, we show that visual and conceptual factors jointly shape the composition of typical scene drawings. By releasing a large dataset of typical scene drawings alongside this work, we further provide a starting point for future studies exploring other critical properties of human drawings.