– E. Vural (EPFL) : Transformation-invariant analysis of visual signals with manifold models

Carte non disponible

Date/heure
Date(s) - 31 janvier 2013

Catégories Pas de Catégories


Transformation-invariant analysis of visual signals with manifold models By Elif Vural, EPFL. Manifold models provide low-dimensional representations that are useful for processing and analyzing visual signals in a transformation-invariant way. The set of images generated by the geometric transformations of a reference visual pattern is called the transformation manifold of that pattern. In an image registration application where the data conforms to a parametrizable geometric transformation model, the registration problem can be formulated as the computation of the projection of the target image onto a reference transformation manifold. Similarly, in a classification application where the data has undergone geometric transformations, the resulting nonlinear structure of data samples can be captured through the use of pattern transformation manifolds. Approximating the data samples of each class with a different transformation manifold, one can estimate the class label of a query data sample by comparing its distance to the candidate class-representative transformation manifolds. In this talk, we will discuss several aspects of transformation-invariant image analysis with manifold models such as the learning and sampling of transformation manifolds.

– E. Vural (EPFL) : Transformation-invariant analysis of visual signals with manifold models

Carte non disponible

Date/heure
Date(s) - 31 janvier 2013

Catégories Pas de Catégories


Transformation-invariant analysis of visual signals with manifold models\n\nBy Elif Vural, EPFL.\n\nManifold models provide low-dimensional representations that are useful for processing and analyzing visual signals in a transformation-invariant way. The set of images generated by the geometric transformations of a reference visual pattern is called the transformation manifold of that pattern. In an image registration application where the data conforms to a parametrizable geometric transformation model, the registration problem can be formulated as the computation of the projection of the target image onto a reference transformation manifold. Similarly, in a classification application where the data has undergone geometric transformations, the resulting nonlinear structure of data samples can be captured through the use of pattern transformation manifolds. Approximating the data samples of each class with a different transformation manifold, one can estimate the class label of a query data sample by comparing its distance to the candidate class-representative transformation manifolds. In this talk, we will discuss several aspects of transformation-invariant image analysis with manifold models such as the learning and sampling of transformation manifolds.[