Combining Corners from Multiple Segmenters

Abstract

Pen-based interfaces utilize sketch recognition in order to allow users to sketch complex systems with intuitive input. In order to allow users to freely draw their ideas without constraints, the low-level techniques involved with sketch recognition must be perfected because poor low-level accuracy can impair a user s interaction experience. Stroke segmentation algorithms often employ single, specific techniques in their attempts to splice strokes into primitives used for visual shape representations. These algorithms each have their strengths and weaknesses, and different segmenters find and miss different corners. We introduce a technique to combine polyline corner results from different segmenters by using a variation offeature subset selection. Our feature subset selection algorithm uses a sequential floating backward selection with a mean-squared error objective function in order to find the best subset of corners. By utilizing our combination method, we were able to achieve all-or-nothing accuracies of 0.926 on polyline stroke data.