A visual comparison of shape descriptors using multi-dimensional scaling

Abstract

In this paper, we investigate the descriptive capabilities of a series of well known shape descriptors, using classical Multi-Dimensional Scaling (MDS). Sammon Mapping is applied to a data-set of six hundred shapes from a real-world multi-component trademark data-set, that have been described using traditional perceptual descriptors, Fourier descriptors and Rosin's descriptors. The maps generated offer considerable insight into how the descriptors encode shape, there is evidence to suggest that traditional perceptual descriptors offer better discriminating performance than Fourier and Rosin's, and also strong indications that an intelligently chosen combination of descriptors could offer increased discriminating capacity over any single descriptor. The mapping techniques discussed in this paper use an arbitrary distance measure, and hence can be easily adapted to map many other forms of image descriptors. This makes them extremely useful for evaluating descriptor performance and also for visual browsing of image databases. © Springer-Verlag Berlin Heidelberg 2003.