Jennifer Weingarten
Rubine Summary


Their paper covered a system (GRANDMA) for creating a gesture recognition
tools and then overview the performance of a tool (GDP) created with the
system. The recognition tool could only interpret one stroke and it
understood gestures through learning on as little as 15 samples. It was
also possible to recognize gestures before they were completed. Jiggle was
removed my removing inputs within three pixels of each other. Then values
were determined for 13 features
•        cos and sin of the initial angle of the gesture
•        length of the angle
•        bounding box diagonal
•        distance between the first and last point
•        cosine and sign of the angle between the first and last point
•        total gesture length
•        total angle traversed
•        sum of the absolute value of the angle at between each mouse point
•        sum of the squared value of the previous angles
•        maximum speed of the gesture
•        duration of the gesture
Then GRANDMA trained on 15 samples to find which weights should be
associated with various features. In general if a gesture was not within
.95  of an existing gesture after the evaluated weighting of the features
it was discarded.

Reasons for choosing features
-must be computable in constant time to the number of input points
-a small change in the input should result in a small change in each feature
-features should have a real world meaning
-there should be enough features to differentiate gestures