Dec. 19 2005 2:30PM

Dec. 19 2005 3:00PM

Geometric invariants provide a convenient tool for handing image distortions caused by changing the viewpoint. These invariants computed directly from image data can be used as discriminating features between object classes in pattern recognition. There are various approaches for constructing invariants, and most of them are based on the theory of algebraic invariants, integral invariants or normalization. Ideally, invariants should be insensitive to rojective transformations, but deriving those invariants is difficult due to the nonlinear nature of the projective transformation, and it is also questionable if such strong invariants are even needed. Instead, translation, rotation, scaling, or eneral affine models are often used as approximations of the actual transformation. Recently, we proposed an affine invariant method called multiscale autoconvolution that is based on a novel approach for deriving geometric invariants. ater, we have also developed other invariants that are based on essentially the same principles. In this presentation, we will first briefly review these invariants, and then introduce a framework that unifies the theory behind the methods. We will show that by processing images in different domains and scales can provide features that are affine invariant and their discrimination power in object classification can outperform most of the prior methods.

International Workshop on Data - Algorithms - Decision Making (Kontaktní osoba: Martin Janžura)