Bibliography:
Wolin, Aaron, Martin Field, and Tracy Hammond. "Combining corners from multiple segmenters." Proceedings of the Eighth Eurographics Symposium on Sketch-Based Interfaces and Modeling. ACM, 2011.
Link:
Summary:
In this paper the authors implement a corner subset selection with corners detected by five algorithms from different papers (ShortStraw, Doublas-Pecker, Paleo, Sergin, and Kim). In their system the stroke finds all the corners detected from the algorithms and combines them into a master set removing duplicates.
The cull this set using a sequential floating backwards selection. It is a greedy algorithm. From this set, they iteratively remove points and measure the error difference between current and previous sets of points. The point removed is the point that makes the least difference in the error metric. Points are floating so they can be added back in. The error is measured as the squared vertical distance from the original stroke to the current set of proposed polylines. To get the preferred set they used the ratio of the difference of errors and determined when this ratio changed the most, i.e. passed a certain threshold.
The threshold was determined from a training set where the number of corners was known. The ran the algorithm and produced collections of strokes for when the ratio of error differences reflected when the number of points overdetermined the corners (false positives) and when the number of points underdetermined the corners (false negatives.) They used the MAD or meadian absolute deviation of each collection to find the dividing point for the threshold, in between the two.
In this paper the authors implement a corner subset selection with corners detected by five algorithms from different papers (ShortStraw, Doublas-Pecker, Paleo, Sergin, and Kim). In their system the stroke finds all the corners detected from the algorithms and combines them into a master set removing duplicates.
The cull this set using a sequential floating backwards selection. It is a greedy algorithm. From this set, they iteratively remove points and measure the error difference between current and previous sets of points. The point removed is the point that makes the least difference in the error metric. Points are floating so they can be added back in. The error is measured as the squared vertical distance from the original stroke to the current set of proposed polylines. To get the preferred set they used the ratio of the difference of errors and determined when this ratio changed the most, i.e. passed a certain threshold.
The threshold was determined from a training set where the number of corners was known. The ran the algorithm and produced collections of strokes for when the ratio of error differences reflected when the number of points overdetermined the corners (false positives) and when the number of points underdetermined the corners (false negatives.) They used the MAD or meadian absolute deviation of each collection to find the dividing point for the threshold, in between the two.
Comments:
I thought this was a very interesting approach to corner finding, using a combination of other corner finding algorithms, and the fact that the all or nothing accuracy was higher than any other algorithm alone.
I thought this was a very interesting approach to corner finding, using a combination of other corner finding algorithms, and the fact that the all or nothing accuracy was higher than any other algorithm alone.
Research Ideas:
No comments:
Post a Comment