Abstract:
In this paper, a new shape descriptor named hierarchical cross ratio contexts for shape matching is proposed to solve the shape recognition problem under projective transformations. First, a coarse-to-fine approach is used hierarchically to calculate the feature for each contour point, which makes the proposed description combine both global geometry and local contextual information. Second, this algorithm modifies the traditional cross ratio, and constructs the invariants using each five points extracted from the shape contour. Finally, dynamic programming is used to compute the similarity between shapes. Experiments demonstrate that our method obtains high recognition performance with relatively low computation complexity and low feature dimension. Moreover, our representation is also open to any other projective invariants, and is easy to extend by combining with other popular descriptors, which well displays their respective advantages.