HP Labs Technical Reports

Exploration of Point Distribution Models in Machine Vision

Freeburn, Catherine

HPL-2000-14

Keyword(s): machine vision; automatic content classification; statistical model building; point distribution models

Abstract: The thesis explores the use of statistical point distribution models in machine vision. The effectiveness of traditional machine vision three- dimensional model building techniques has been limited, particularly when the objects being modelled are complex, smoothly deforming bodies such as human beings. Recently, a new approach to model building has been found that allows effective two-dimensional image space models of complex non-rigid objects to be automatically generated by statistically analysing a set of training images. One such technique creates what is known as a point distribution model.In this thesis I explore whether point distribution models still perform well in areas where the objects being modelled are much less complicated than have hitherto been tried. This was achieved by creating a rational reconstruction of an image interpretation system that automatically generated and used a point distribution model to recognize human beings, but to then use the resulting system to build (and use) models of much more simple rigid objects: specifically, convex polygonal prisms. The system that was reconstructed was a pedestrian tracking system designed and implemented by Baumberg, as part of a Leeds/Reading university collaboration project. It was found that for simple convex prismatic objects Bamberg's mechanism for automatically extracting point descriptions of object silhouettes (which are then used to train a model) fails to produce outline descriptions in which the positions of the points move smoothly as the silhouette of the objects change. This problem meant that the resulting models were unstable and in the experiments done could only correctly classify objects just under 50% of the time. In addition, early indications were found that suggest that even if a more stable outline extraction method can be found, there is unlikely to be enough variation in the outlines of simple objects to enable this kind of model to be able to systematically distinguish between them. Notes: M.S.c Project Report

69 Pages

Back to Index