Edge points sampling in shape context

Question

I'm recently preparing to use shape context and SVM to do character recognition. The histograms in shape context are considered as feature vectors of characters. However, after edge detection, how to sample the edge points of two characters' edges that may have different points totally? Sampling randomly seems not good.

Answer 1

Shape context uses a fixed number of points for every shape it analyses, so you have to filter your edge detection results and drop unnecessary points.

If the edge detection does not give you enough points you should consider lowering your number of points used for the shape context in total. But there should also be a fallback which interpolates points if needed. Linear interpolation is good enough for a fallback.

The points used should be more or less evenly distributed along the edge of the shape. If the points come in large clusters shape context will not work at all (there are problems with the normalization of the radial distances required by the scale invariance too)