A domain-independent system for sketch recognition

Bo Yu and Shijie Cai. 2003. A domain-independent system for sketch recognition. In Proceedings of the 1st international conference on Computer graphics and interactive techniques in Australasia and South East Asia (GRAPHITE '03). ACM, New York, NY, USA, 141-146.


This paper examines techniques for a domain independent sketch recognition system. The approach can be divided into two steps: imprecise stroke approximation and post-process. The first step is applied as the user sketches, while the second is applied after the entire drawing is completed. The output of the system is a hierarchy which includes raw stroke data, vertex and primitive shape data, and semantic information.

For stroke approximation, the system uses direction and curvature coupled with feature area. Vertex detection and primitive shape approximation are combined into a single incremental procedure. A recursive approach is first applied, splitting a stroke until its pieces are recognized as primitives. Lines are approximated by fitting a line to the stroke points or the curve graph. Feature area is then used to judge validity. If the ratio between feature area and the candidate line's area is less than 1.0, a line is detected otherwise an arc detector is applied. To detect a circle or ellipse the direction graph should be able to be fitted with a line of slope close to 2PI / n, where n is the total number of stroke points. For arcs, the slope of the line fit to the direction curve should be less t han 2PI/n.

The paper also addresses self intersection strokes. If given a stroke that has self intersection, two copies are made which each follow a different division strategy (one intersection point, one maximum curvature point). The recognition results are then compared and the better is chosen.

The second phase, post-process eliminates false elements, adjust the size and position of elements, and performs object recognition.

This paper contributes some novel techniques, specifically feature area, for performing domain independent recognition of shapes. However, it seems to only work for simple shape drawings. It would be interesting to see this applied to more complex examples.