Constructing Spatial Representations of Variable Detail for Sketch Recognition

摘要

We describe a system which constructs spatial representations of sketches drawn by users. These representations are currently being used as the input for a spatial reasoning system which learns classifiers for performing sketch recognition. The spatial reasoning system requires representations at a level of detail sparser than that which the representation constructor normally builds. Therefore, we describe how the representation constructor ranks the expressions in its output so that the number of expressions in the representation can be decreased with minimal loss of information. We evaluate the overall system, showing that it is able to learn and utilize classifiers for complex sketches even when the representation size is sharply diminished.