Application of Sparse Gaussian Process in Visual Object Category Recognition

Abstract

Gaussian Process as a non-parametric regression technique has witnessed renewed interest in the past couple of years. In this thesis we provide a broad overview of GP, and try to apply the recent advancements to our problem. We target the problem of visual object category recognition using Gaussian Processes Regression in this proposal. The focus is primarily on the design of localized combination of kernel functions for visual object categorization, as well on the sparse approximations to Gaussian Processes. We use a linearly weighted kernel function combination to shape the smoothness of the joint GP prior. These kernel weights are determined using evidence maximization in some sample space of the training data. This sample space is determined by the test image in question. The kernel weights are sensitive to the test image in question, and are local in nature, as opposed to global weights that would be same for all test images. We incorporate sparse inferencing techniques in order to counter the cubic time complexity of Gaussian processes inferencing. Also we briefly explore the application of Gaussian Process Regression in the domain of image denoising. A simple GP based denoising scheme is implemented to demonstrate the effectiveness of GP in such applications.