Abstract
Line drawings can be an effective way of conveying the 3D shape of an object. Computational models for line drawing interpretation have mainly dealt with “blocks world” drawings, with clean lines corresponding to the sharp creases between faces. However, humans can interpret drawings of blobby, “organic” objects, where there are no well defined faces, and where the lines do not lie on creases. Moreover, humans can deal with lines that are broken, noisy, and and disconnected. We have developed a computational model that mimics human performance on this kind of imagery. To produce a drawing interpretation, the system uses a novel combination of techniques from object recognition and 3D shape modeling: a machine learning stage first estimates the figure-ground direction of each line, and a shape optimization stage then finds a smooth surface that satisfies the figure-ground constraints. The system requires no initial labeling or processing by a human. We compare the results of our system to the surface orientations reported by humans when shown computer-generated drawings of known 3D shapes. We can thus compare our results both against the aggregate human perception of each drawing, and against the original 3D shape from which the drawing was made. For the restricted class of smooth shapes, we find that the algorithm produces very similar interpretations to humans, both where the humans interpret the original 3D shape accurately and where they do not. This success suggests that the strategies used by the algorithm may be similar to those employed by humans.