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

Authors : Brandon Paulson, Pankaj Rajan, Pedro Davalos, Ricardo Gutierrez-Osuna, Tracy Hammond

Comments:
1. Andrew's blog
Summary:
This paper constructs a sketch recognition system with 44 features - 31 from paleo-sketch and 13 from rubine and uses a quadratic classifier to recognize sketches. The results are then compared to Paleo sketch. The samples used to train and test the system where the same as that used for paleosketch.
Of the 44 features , the system identifies the following features as important through feature subset selection:

* End point to stroke length ratio -
* NDDE & DCR
* Total rotation - above 4 features were used more than 90% to classify

* Curve least square error - to identify curves

* Circle fit : Major axis: minor axis - used to classify circle/ ellipse

* Spiral fit - Average radius to bounding box radius ratio, Center closeness error

* Polyline fit - # of substrokes, number of strokes passing line test, feature area error

* Complex fit- # of sub-fits, # of non polyline primitives, percent of subfits that are lines, complex score/ rank

Results show this system produces 97% accuracy on this optimal subset which is as good as the paleo sketch system(98.56%)

Discussion:
This is a very interesting set. Each feature is significant to identify a particular type of primitive.
I do not understand how the system classifies the arc. I think it considers arcs as subset of curves.
Its interesting to see the use of Polyline fit: feature area error, i think this helps the system to distinguish between ellipse/circle and polylines.

Calculating percent of substroke that are lines for Polyline and complex fit seems to be rendundant. But both of them are marked important by this system.

For curves , least squares error is significant and for poly lines, its feature area which is important. its quite different from what we have discussed in the class. I thought least squares would be more significant for lines and feature area for curves.