A New Approach to Early Sketch Processing

Early processing of sketches is a requirement common to many pen based systems. The task is difficult because of variations between sketches from different users and in different domains, and because of ambiguities that arise within even slightly messy sketches, especially when the input is allowed to be unconstrained. We propose a graphical model based approach to early sketch processing. Small areas of a sketch corresponding to features such as corners and straight segments are considered individually, and a likely labeling for such features is found by incorporating some context in order to improve on labels computed with only local information. Results from applying this approach to the problem of detecting corners show an improvement.