Australian Centre for Visual Technologies Blog

Chunhua Shen Talk
February 5, 2010, 5:25 am
Filed under: events

On 12 Feb, Dr Chunhua Shen from NICTA will be presenting a talk on ‘A Duality View of Boosting Algorithms’:

Consideration of the primal and dual problems together leads to important new insights into the characteristics of boosting algorithms. We show that the Lagrange dual problems of AdaBoost, LogitBoost and soft-margin LPBoost with generalized hinge loss are all entropy maximization problems.

By looking at the dual problems of these boosting algorithms, we show that the success of boosting algorithms can be understood in terms of maintaining a better margin distribution by maximizing margins and at the same time controlling the margin variance. We also theoretically prove that, approximately, AdaBoost maximizes the average margin, instead of the minimum margin. The duality formulation also enables us to develop column generation based optimization algorithms, which are totally corrective.

Based on these new theoretical results, a new boosting algorithm is designed. We call it margin-distribution boosting (MDBoost). MDBoost directly maximizes the average margin and minimizes the margin variance simultaneously. Empirically we show it outperforms AdaBoost and LPBoost on the UCI machine learning data sets.


Peter Kovesi seminar
December 11, 2009, 1:42 am
Filed under: events

On 27 Nov, Peter Kovesi from the University of Western Australia presented a seminar on Animating Impossible Objects:

For some impossible objects it is possible to construct a three-dimensional model that, when viewed from a particular direction, gives rise to the impression that the object is impossible.  The impossible triangle is one such object.  These three-dimensional models of impossible objects can only be viewed from one angle – otherwise they no longer look impossible. But is it possible to create an interactive impossible object, that is, an impossible object that can be viewed from any angle? This talk explores the creation of such objects on the computer.

To allow an impossible object to be viewed from any angle its 3D geometry must be altered to suit the viewpoint. We show that a particular class of impossible figures can be described in terms of two complementary halves. The complementary halves are related to each other by an inversion transformation in the image plane.  Either one of the complementary halves can be realized as a 3D object with the appropriate geometry to ensure an impossible figure will be produced.  The 3D model of one of the complementary halves can be animated by normal means. Once an arbitrary view has been generated, the other complementary half can then be constructed by image plane inversion to complete the impossible figure.

Michael Brown seminar
December 11, 2009, 1:36 am
Filed under: events

On 26 Nov, Dr Michael Brown from the National University of Singapore presented a seminar at the ACVT titled ‘Interactive computer vision: Exploiting meaningful interaction in computer vision applications’:

A dark secret in the computer vision and image processing research community is the heavy reliance on magic numbers by our algorithms. Magic numbers often manifest themselves as algorithmic parameters that must be tuned before satisfactory results can be obtained. This over reliance on parameter tuning stems, in part, from the long-standing dogma that computer vision algorithms should be fully automated. Interestingly, requiring the user to tune parameters, most of which have no intuitive meaning to the task at hand, is far from automatic. In fact, it is a major stumbling block when building real world computer vision applications.

In this talk, I will advocate that for many computer vision and image processing applications magic numbers can be avoided if we instead exploit the user’s help via meaningful interaction. This approach to solving problems has been termed interactive computer vision and has proven effective in many tasks such as segmentation, matting, and image repair. Specifically, I discuss several examples from our own research that have transformed problems either too difficult to automate or heavily reliant on parameter tuning into applications that now rely only on simple, and easy to understand, interaction supplied by the user.