An Interpolation Based Approach For Pattern Recognition And Generation

Abstract

A large number of problems in computer vision and computer graphics can essentially be reduced to a pattern recognition problem. In this thesis, we explore a novel interpolation based framework to address some of the various recognition problems in these areas. Our interpolation based framework is a supervised learning algorithm that allows for both generation (synthesis) of new patterns as well as perception (analysis) of existing patterns. The method is simple to implement and yet, expects a very straightforward and intuitive set of parameters to model the complex nature of such recognition problems.Specifically, given a set of training data along with their parameters, we can learn a model that is a compact representation of the set of all patterns defined in a parametric space. Having learnt such a model we are able to generate any new patterns defined within that parametric space. Moreover, as an inverse operation, we are also able to estimate the parameters of any existing pattern. Based on this 'synthesis-analysis' approach we propose a method to recognize patterns and evaluate it in rather diverse areas such as recognition of objects/faces in varying illumination conditions and, human motion across different skeleton sizes. Using the same approach we demonstrate the methods application in the area of image based modeling and rendering, where, we are able to render `unknown' objects into a scene provided we have at least one `known' object in it. Another application is in the area of animation where, given a set of human motion data differing in skeleton size but for a specific action, we are able to re-target that specific action to an identical skeleton but of varying bone lengths.Also, in this thesis, we explore a novel image feature descriptor built using a bank of Gabor filters and evaluate its effectiveness in an object recognition framework using synthetic and real data. We also describe our software tool that allows for automatic generation of ground-truth data for various computer vision problems such as camera calibration, feature matching, 3D reconstruction, object tracking and object recognition.