Reading 11 : Combining geometric and gesture-based features

Citation:
Paulson, Brandon, et al. "What!?! no Rubine features?: using geometric-based features to produce normalized confidence values for sketch recognition." HCC Workshop: Sketch Tools for Diagramming. 2008.

Publication Link

Summary:

The main aim of this paper is to investigate a combination of geometric features (how it looks), and gesture based techniques to build a highly accurate classifier, and produce normalized confidence values. The authors than proceed to finally establish how gesture based features turn out to be insignificant when an optimal set of features is computed using a greedy algorithm to maximize accuracy. 

Discussion:

Geometric based recognisers focus on determining the error between a sketched shape and its ideal version using a series of geometric tests and formulas. A total of 44 features were selected for evaluation (31 geometric, 13 Rubine's), and a dataset consisted of 90 samples each from 20 users was aggregated.

The relevant features are determined using a greedy, sequential forward selection algorithm. Optimal features are the ones that that yield highest accuracy during the SFS (Subset feature selection) process. Using a 50-50 split (same as PaleoSketch), the quadratic classifier was found to perform poorly when the entire set of features was used. When a an optimal set of features (obtained thorough multi-fold validation) was used, the quadratic classifier was able to match Paleosketch. 

The advantages of this system over Paleosketch is that it is easier to code, and is computationally faster since it makes of an optimal subset of features. Since the data is partitioned userwise, the subset selection algorithm tends to bias towards user independence, and most gesture-based features were found to be insignificant.