Dae Hyun Kim and Myoung-Jun Kim, A curvature estimation for pen input segmentation in sketch-based modeling
Kim & Kim's paper takes somewhat of a middle ground between Sezgin and Yu. While Sezgin's paper relied equally on speed and curvature to detect the end of stroke subsegments, and Yu relied only on curvature, the Kims only uses speed at specific points to help aid the choice of curvature points. They also present two new methods for computing curvature. The first incorporates curvature of adjacent points to not only capture how curvy the line is a specific point, but also in the surrounding regions. To do this they simply sum up the curvature values over a window centered at the point in question if the values have the same sign as the curvature at that point. The second method imposes a monotonicity constraint on the window as well, considering only curvatures of the same sign with a smaller magnitude.
The speed-adaptive threshold is an interesting compromise between Sezgin and Yu, seemingly making it more likely that a curvature corner will be found at a lower speed. Broadening the window for which we examine curvature does seem improve the recognition of corners, but at the same time it may smooth out intentional, small changes in the stroke.
Subscribe to:
Post Comments (Atom)
2 comments:
I would be interested to see how broadening or shortening the window would work with this algorithm, since the number of points in the window might not matter as much. The curvature cutoffs (based on direction sign changes and value shifts) could allow for more points within the window without causing too much smoothing.
I'd be interested in selective smoothing. What are the primary types of speed/curvature or other feature values that signal an area that would benefit from it?
Post a Comment