Viirya

"Shape matching and object recognition using shape contexts"

By Serge Belongie, Jitendra Malik and Jan Puzicha

--

The authors in this work try to operationalize the notion of shape similarity.

The way they approach this problem is to find correspondences between shapes.

First they represent shapes as a set of points sampled from shape contours. They introduce a shape descriptor, shape context, to describe a given point on a shape by the coarse distribution of the rest of the shape.

Finding correspondences between two shapes is equivalent to finding the point that has the most similar shape context on one shape with respect to each point on the other shape.

The robust and simple algorithm for finding correspondences between shapes is their main contribution in this work.

In order to extend the correspondences to the complete shape, they estimate an aligning transformation that maps one shape to the other.

The authors define the dissimilarity between two shapes as a sum of matching errors between corresponding points with a term used to measure the magnitude of the aligning transform.

Given the dissimilarity measure, the problem of object recognition is solved using nearest-neighbor techniques.