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

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

This paper explores the merging of geometric recognition techniques and gesture-based recognition techniques into a single recognizer. The recognizer produced is capable of allowing natural sketches to be classified, while offering normalized confidence values for alternative interpretations. The major surprising result is that geometric features are more helpful for recognition than gesture-based features, when given naturally sketched data.

The paper utilizes a statistical classifier, a quadratic classifier, for examining features from both geometric and gesture space. Initially 44 features were used (31 geometric, and 13 Rubine). Testing was conducted using a 50% split of the data, split by user, so that the system was not trained on data from the participant whose sketches it would be attempting to classify. Feature Subset Selection was employed to order the entry of features. Using only the features which were present at least 50% of the time, the system is able to achieve results that are not statistically significantly different from PaleoSketch.

There are several interesting aspects to this paper. First, it is noted that with only the top six features, 93% accuracy is still achievable. Additionally, only one gesture-based feature, total rotation, was chosen as one of the top features. This is interesting because the testing setup was more typical of a real world use system, where the system is not trained with the user, but is independent between users. This provides further evidence that geometric based properties are more useful for user independent designs.