Abstract
Why is it that we can recognize objects and understand 3D shapes from line drawings, even though line drawings do not exist in the natural world? One theory is that we learn line drawing perception, just as we learn written language (Goodman 1968). Yet, even observers who have never seen pictures before can recognize objects in line drawings, e.g., Kennedy and Ross (1975), Jahoda et al. (1977). According to a more recent theory (Sayim and Cavanagh 2011), line drawings activates edge receptors in V1 in order to produce the same percepts as corresponding natural images. However, this theory does not account for the many differences between edge images and natural images, e.g., natural images often have different edges from line drawings. We hypothesize that the visual system perceives a line drawing as though it were a realistic image under a specific type of lighting and material conditions. This hypothesis is based on Abstracted Shading (DeCarlo 2003, Lee 2007, Pearson and Robertson 1985), which shows how plausibly realistic conditions can produce drawing-like images, and is known to be predictive of hand-made drawings (Cole 2008). In other words, basic line drawing styles work because they are visually similar to plausibly-realistic images. This similarity applies to all visual features, not just edges. More generally, we hypothesize that basic line drawing perception is a consequence of realistic image perception, independent of the perceptual mechanism (e.g., edge receptors). We test this by showing that a computer vision model trained only on real images can infer depth from line drawings. These hypotheses suggest new ways to understand line drawing perception as a special case of natural image perception.