An Eigenvector Approach Based on Shape Context Patterns for Point Matching

In this paper, the problem of point correspondence across two images is treated in the eigenvector analysis matching framework of Scott and Longuet-Higgins. We develop the concept of shape contexts introduced by S. Belongie et al. to shape context patterns as rich local descriptors of points. We further propose a Gaussian-weighted Hausdorff distance between shape context patterns to measure correspondence strength in Scott and Longuet-Higgins framework. The resultant point matching approach is applied to estimate affine transformation between handwritten Chinese character images, whose effectiveness is confirmed by the experimental results