Meaningful Matches: Experiments on LLD and MSER

This chapter tests the shape matching method described in the previous chapter. Section 6.1 deals with the semi-local invariant recognition method. Both similarity and affine methods are considered, and a comparative study based on examples is presented. When images differ by a similarity, affine matching usually returns less matches because affine encoding is more demanding. Nevertheless, affine encoding proves more robust as soon as there is a slight perspective effect, and yields much smaller NFAs.We will also test an improved MSER method (namely a global affine matching algorithm of closed level lines). This algorithm works but we will point out a problem with convex shapes, which turn out to be very hard to distinguish up to an affine transformation. Finally the context-dependence of recognition will be illustrated by striking experiments on character recognition.

Now comes the time to check the applicability of the shape comparison scheme described in the previous chapters. All the experiments presented thereafter follow the same procedure: detection of meaningful boundaries (Chap. 2), affine invariant smoothing (Chap. 3, Sect. 3.3), similarity or affine normalization-encoding (Chap. 3 and 4), and then matching (Chap. 5).