Aligning shapes for symbol classification and retrieval

This paper proposes a method able to exploit peculiarities of both, local and global shape descriptors, to be employed for shape classification and retrieval. In the proposed framework, the silhouettes of symbols are firstly described through Bags of Shape Contexts. The shape signature is then used to solve the correspondence problem between points of two shapes. The obtained correspondences are employed to recover the geometric transformations between the shape to be classified/retrieved and the ones belonging to the training dataset. The alignment is based on a voting procedure in the parameter space of the model considered to recover the geometric transformation. The aligned shapes are finally described with the Blurred Shape Model descriptor for classification and retrieval purposes. Experimental results demonstrate the effectiveness of the proposed solution on two classic benchmark shape datasets, as well as on a large scale set of hand sketches composed by 20,000 examples distributed over 250 object categories.