Vectorisation and interpretation of drawings with artistic cues

Abstract

Computer modelling from two dimensional sketches is becoming more popular among people who wish to engage in modern design. Although the interpretation of sketches seems to be a trivial problem for human observers who have an innate ability to see and understand depth in drawings, machine interpretation of sketches is a nontrivial problem and has been the subject of research, particularly since a simple drawing may have a large number of possible 3D interpretations, of which, only a select few are deemed plausible by human observers. While edge boundary strokes are sufficient to elicit a 3D interpretation of the sketched object, humans do not rely solely on the edge boundary strokes and often introduce additional artistic cues to aid other observers into interpreting the sketch. Such artistic cues introduce additional difficulties in the separation of the object sketches from the artistic cue. To our knowledge, shadows have only been used in the interpretation of edges from photographic scenes where labelling algorithms are used to distinguish shadow edges from object edges rather than to modulate the 3D interpretation of the object. Our contributions to this research area are twofold. We first describe, implement and evaluate a circle-based vectorisation algorithm that allows us to obtain vector representations of sketched objects even when these contain shadows. This vectorisation algorithm uses evidence accumulated from concentric sampling circles which allows us to detect junctions and lines even when these are off-centered from the circle samplers, hence extending the current state-of-the-art in drawing vectorisation algorithms which require clean images and centred samplers. We further observe the artistic cues that are used by artists in sketches and develop an edge labelling interpretation algorithm that uses shading and table line cues to modulate the interpretation of the sketch according to the design intent as portrayed by the cues present in the sketch. Since human observers can tolerate a degree of error in the sketched cues, we cast the edge interpretation algorithm in a genetic algorithm framework, thereby allowing our interpretation algorithms to determine edge interpretations when cues are missing or drawn incorrectly. We evaluate our algorithms on sample drawings created by nine volunteering participants, obtaining a vector representation and subsequently an edge interpretation of the drawings that corresponds to the drawing intent.