What no Rubine Features? Using Geometric-based Features to Produce Normalized Confidence Values for Sketch Recognition

Comments

Daniel's blog

Summary

In this paper the author tries to merge the two techniques of Geometric-based sketch recognition and Gesture-based sketch recognition. Both the techniques have their advantages and disadvantages and the author's focus is to get the best out of both the techniques.

For this the author modified the Rubine algorithm to use a quadratic classifier instead of a linear classifier. He integrated all the 31 features of the PaleoSketch and also 13 features from the Rubine algorithm. This called the full feature set. With the full feature the accuracy rate was not comparable to the best accuracy available from the PaleoSketch.

The quadratic classifier can be optimized by removing the features which had the negative effect on recognition. For this author tried to pick out a subset of features which could give the best result. The technique used was the sequential forward selection (SFS). In which the author tried all the possible combination starting with a feature set of 1 feature. The author achieved better accuracies with the optimal feature set.

Interesting to note was that a 93% accuracy was achieved with only top six features. Only one of the Rubine feature was which is total rotation was included in the top six features.

Discussion

The author uses a different approach by merging the feature sets of the two different recognizers to produce a feature set which is most significant in the recognition of the sketch. The accuracy achieved is nearly equal to the hand tuned algorithm PaleoSketch. As it uses a classifier more prmitives can be easily added to the system and easily trained for the recognition of the new primitive.

The only problem that i see is that its application is only tested on a very limited set of symbols and there is high probability that its result wont be as good for complex shapes.