This thesis presents methods for improving sketch understanding, without knowledge of a domain or the particular symbols being used in the sketch, by recognizing common sketch primitives. We address two issues that complicate recognition in its early stages. The first is imprecision and inconsistencies within a single sketch or between sketches by the same person. This problem is addressed with a graphical model approach that incorporates limited knowledge of the surrounding area in the sketch to better decide the intended meaning of a small piece of the sketch. The second problem, that of variation among sketches with different authors, is addressed by forming groups from the authors in training set. We apply these methods to the problem of finding corners, a common sketch primitive, and describe how this can translate into better recognition of entire sketches. We also describe the collection of a data set of sketches.
展开▼
