shape context descriptors are then indexed by a locality-sensitive hashing data structure aiming to perform approximate k-NN search in high dimensional spaces in sub-linear time.
>>>Efficient Logo Retrieval Through Hashing Shape Context Descriptors.
M. Rusiñol and J. Lladós. In Proceedings of the Ninth IAPR International Workshop on Document Analysis Systems, DAS10, pages 215-222, 2010. [pdf]
CSC – color shape context
For each sampled point pi, shape context histogram Si, To add color info to SC:
- A circular mask is defined as the region of interest centered at pi, the size of the mask is computed with relation to the mean distance between all the points pairs in the shape.
- The local color name histogram Ci
- combination of both descriptors Si and Ci at each point of the shape –
- Distance between two point d(pi, pj):multiplying the distance of the shape context descriptor and the local color descriptor
- Matching between sketch&image: Given a set of local distances d(pi, pj) between all pairs of points, the
final distance between the sketch query and the image is determined by minimizing the total cost of matching
bipartite graph matching approach that puts in correspondence points having similar shape and color descriptions.
- Once all the n points in a shape are described by their shape context histogram, in order to match two shapes we have to find the point correspondences.
- The simplest way to compute the matching among the two set of points is by using a bipartite graph matching approach
- In order to obtain a more robust matching, the most usual techniques involve the computation of an affine transform that matches the set of points from one shape to another.
>>>>Perceptual Image Retrieval by Adding Color Information to the Shape Context DescriptorM. Rusinol, F. Nourbakhsh, D. Karatzas, E. Valveny and J. Llados Proceedings of the 20th International Conference on Pattern Recognition, IEEE Press, pp. 1594-1597, Istanbul, Turkey, 2010 [pdf]
shapeme histogram descriptor was proposed by Mori et al. 
This description was inspired by the bag-of-words model. The main idea is to do vector quantization in the space of the shape context of all interest points.
- clustering stage of the shape context feature space: k-means algorithm
each shape context descriptor can be identified
by the index of the cluster which it belongs to
- extract n sampled point from edge map, then get the shape context hi
- then each shape context descriptor of the point is projected to the clustered space identified by single index Ii
- the query image can be represented by a histogram coding the frequency of appearance of each of k shapeme indices
- to find the matches, is to find the kNN in the spaece of shapeme histogram.