Metric similarities learning through examples: An application to shape retrieval

Abstract

The design of good features and good similarity measures between features plays a central role in any retrieval system. The use of metric similarities (i.e. comingfrom a real distance) is also very important to allow fast retrieval on large databases. Moreover, these similarity functions should be flexible enough to be tuned to fit users behaviour. These two constraints, flexibility and metricity are generally difficult to fulfill. Our contribution is two folds: We show that the kernel approach introduced by Vapnik, can be used to generate metric similarities, especially for the difficult case of planar shapes (invariant to rotation and scaling). Moreover, we show that much more flexibility can be added by non-rigid deformation of the induced feature space. Defining an adequate Bayesian users model, we describe an estimation procedure based on the maximisation of the underlyinglog -likehood function.