Analysis of “A curvature estimation for pen input segmentation in sketch-based modeling”
Comments Made Elsewhere:
Summary:
Paper summaries research on discovering the “region of support” (number of neighbors) to look at when determining curvature on the on-the-fly and only using angle space.
Things to note: resampling will undoubtedly cause smoothing of the input curve, so determining a good size for the k-neighbors is necessary.
As a first step, they initially define the curvature as the direction change at each point with k = 1 (only one neighbor). Next, they define “local convexity” where all neighboring points with the same direction (sign) as point(i) are used as the region of support.� Works well for some situations, but came make features indistinguishable (where to segment?).� Added “local monotonicity” where you only observe the neighboring subsets on the left and right as long as they are monotonically decreasing (meaning keep decreasing).� Next, find the local maximum or minimum (depending on if curve is positive or negative) to find the segmentation point.
Performed evaluation test with volunteers and then tested their algorithm in parts (direction only, with local convexity, only, etc.) and then against two other accepted algorithms.
Discussion:
Again, another paper that doesn’t make use of time so their implementation is independent of input medium (tablet PC or actual paper) and the direction of the scanning (can start from the front or back). They do use speed for calculating the threshold to use when determining the segmentation points.
I was actually thinking up a solution similar to theirs (but theirs is obviously better).� I did not realize the issue of determining the optimal number (k) of nearest numbers to use, but am supportive of using a sliding window at drawing time for calculating the segmentation points there and then.
I like that they do use their segmentation algorithm for going to the next step of defining features for shape recognition on page 5.� This algorithm appears to be capable of complex line segmentation.
How could this be done without resampling?� Is it possible in determining curve segmentation?� I want to develop a vector-based approach that incorporates speed and doesn’t require resampling.
[...] my discussion in the last blog post, here’s a solution for corner finding that doesn’t appear to require resampling. It [...]
the sliding window works better than feature calculation for a single point, as it reduces the errors due to jitters. I would like to know in detail about you approach.Could you explain that? Almost all the corner finding algorithms which we learned so far have failed for curves with irregular curvatures. Does your algorithm work for these cases?