Analysis of “Ink Features for Diagram Recognition”
Comments Made Elsewhere:
Summary:
Claims not much has been done on determining with features that are important, and on determining text versus shape. Claims that recognition for text and shape must be handled different.
Created a list of 40+ features, generated and hand labeled a corpus of strokes, then did statistical analysis to create a binary classification tree of eight features. Found inter-stroke gaps (time between each stoke) to be important feature.
Discussion:
This is still single stroke recognition, if I’m not mistaken (i.e. each stoke is either a shape or letter).
Simple to understand, yet inflexible as they mentioned (they also suggest a HMM). Kind of half-gesture, half-geometric features. Wonder how computer vision could be used here (not really wanting to just recognition, just be able to put the shape down the tube of the right recognizer). Not sure it could since relying on how it was drawn instead of what was drawn avoids training and misclassification of shapes that accidentally look like letters. Using time and size seeks to capture more of the user’s intent.
Not sure why they showed the results from running the dividers on the training sets (Figure 4). I guess just to remphasize how bad the other two are.
One theory: text will have more cusps.
Interesting theory… But Cusps will be helpful only for the cursive writings. Capital letters makes things difficult for us in this case. Capital letters are usually written in more than one strokes.