Bibliography:
Tevfik Metin Sezgin, Thomas Stahovich, and Randall Davis. Sketch Based Interfaces: Early Processing for Sketch Understanding. Proceedings of the 2001 Workshop on Perceptive User Interfaces. pp. 1-8. 2001.
Summary:
The paper describes a method for early stage processing of sketches. The motivation of early stage processing of sketches is to make sketching a natural way to comunicate with computers, without having to deal with icons, menus and tool selection. The system consists of three phases 1) Approximation 2) Beautification 3) Recognition.
Approximation is done by first finding vertices, then detecting curved segments of the stroke. Vertices (or corners) are found by using the direction curvature and speed data of the strokes. Extrema are curvature and speed are regarded as corners, only after reliably removing extrema due to noise. This is done by using average based filtering. This is done by looking at values beyond a threshold. For curvature, max values are found only above a certain threshold. Similarly for speed, min values are found only below a certain threshold. A hybrid fit (of both speed and curvature) is used to determine vertices, from a final set of candidates. (which were obtained from both speed and curvature extremas)
In order to detect curves, bezier curves are fit to the polyline fits created by the hybrid models for detecting corners. For computational efficiency, we use a piecewise linear fit on the bezier curve and compute the error for this approximation. If this error is above a given error-tolerance, the curve is subdivided in the middle and the process is repeated.
After approximation, beautification is done to the result, to make it look as intended. Slopes of a cluster of line segments are weighted and averaged. Once beautification is done, basic objects are recognized from the low-level line segments and curves. This is done by template matching for simple figures.
The system was tested with 13 users, and had a high accuracy of 96 percent. Most users preferred using the system's interface over paper and tablets.
Discussion:
Identification of corners and curves (at corners) was something I was taking for granted in sketch recognition systems. The idea of using speed and curvature data for identifying vertices is pretty cool. I wonder how these methods compare to corner identification techniques from the computer vision domain. Is there a real advantage to using sketching information rather than just using imaging recognition techniques to do the same?
Tevfik Metin Sezgin, Thomas Stahovich, and Randall Davis. Sketch Based Interfaces: Early Processing for Sketch Understanding. Proceedings of the 2001 Workshop on Perceptive User Interfaces. pp. 1-8. 2001.
Summary:
The paper describes a method for early stage processing of sketches. The motivation of early stage processing of sketches is to make sketching a natural way to comunicate with computers, without having to deal with icons, menus and tool selection. The system consists of three phases 1) Approximation 2) Beautification 3) Recognition.
Approximation is done by first finding vertices, then detecting curved segments of the stroke. Vertices (or corners) are found by using the direction curvature and speed data of the strokes. Extrema are curvature and speed are regarded as corners, only after reliably removing extrema due to noise. This is done by using average based filtering. This is done by looking at values beyond a threshold. For curvature, max values are found only above a certain threshold. Similarly for speed, min values are found only below a certain threshold. A hybrid fit (of both speed and curvature) is used to determine vertices, from a final set of candidates. (which were obtained from both speed and curvature extremas)
In order to detect curves, bezier curves are fit to the polyline fits created by the hybrid models for detecting corners. For computational efficiency, we use a piecewise linear fit on the bezier curve and compute the error for this approximation. If this error is above a given error-tolerance, the curve is subdivided in the middle and the process is repeated.
After approximation, beautification is done to the result, to make it look as intended. Slopes of a cluster of line segments are weighted and averaged. Once beautification is done, basic objects are recognized from the low-level line segments and curves. This is done by template matching for simple figures.
The system was tested with 13 users, and had a high accuracy of 96 percent. Most users preferred using the system's interface over paper and tablets.
Discussion:
Identification of corners and curves (at corners) was something I was taking for granted in sketch recognition systems. The idea of using speed and curvature data for identifying vertices is pretty cool. I wonder how these methods compare to corner identification techniques from the computer vision domain. Is there a real advantage to using sketching information rather than just using imaging recognition techniques to do the same?
No comments:
Post a Comment