"GLADDER: Combining Gesture and Geometric Sketch Recognition"

by Paul Corey and Tracy Hammond

Summary

This paper discusses an approach to combine Rubine features with the geometric classifier of LADDER.

Modified Rubine Recognizer: In addition to average feature vectors, class specific covariance matrices are maintained for each class. Input strokes are assigned to a specific class based on computation of Mahalanobis distances.

LADDER Recognizer: Primitives are recognized by low-level shape tests, and are orders based on the hierarchical classifier. Top three are chosen as possible interpretation, and less complex interpretations are put before more complex ones.

Combination: The minimal Mahalanobis distance of a stroke to any Rubine class is computed, if it exceed a threshold of 35, then it's likely to be a LADDER primitive (average Mahalanobis distance 100), rather than one of the Rubine glyphs (average Mahalanobis distance 24). If it is below the threshold, then Rubine interpretation is put in top of the fit list, otherwise, Rubine interpretation is put at the bottom. Context can later be used to rectify an incorrect ordering of fits.

The accuracies are: Modified Rubine 61.9%, LADDER 75.2%, Integrated 79.9%. Some errors are due to overlaping of classes between Rubin and LADDER (zero and circle). And they also suggests in future work that a tiered thresholding system can be used to put Rubine glyphs after less complex primitives and before complex primitives of LADDER interpretations.


Discussion

The combination approach partly solved the problem of the mix of characters and shapes in a sketch. It's a step forward the "free" or "natural" sketch recognition. The tiered thresholding system may be implemented by dividing LADDER primitives and Rubine classes into groups, whose average Mahalanobis distance are quite different, but the deviation of each class within the group is small.