Neighborhood Graph Construction for Semi-Supervised Classification
Neighborhood graphs are used to model the data manifold in semi-supervised classification tasks consistent with the manifold assumption. Since the manifold assumption expression is heavily relied on the manifold model, neighborhood graph construction plays a key role in the quality of the classification for such problems. In this project we aim to develop solutions for this problem that are both robust against the noise in the data, and adaptive to the classification task.The proposed methods are evaluated using the standard UCI datasets that are compatible with the manifold assumption. These may include MNIST and USPS for digit recognition, ISOLET for spoken letter recognition, Newsgroup for text classification and Covertype for forest cover type classification.
Members : Mohammad Hossein Rohban
Video Error Concealment Using the Gaussian Process Framework
In this project we will use the a Gaussian Process to model the video information. Therefore, the problem of estimation of the lost blocks in a video frame reduces to the Bayesian estimation of Gaussian Process given the process values in the correctly received blocks. The problem of designing the appropriate covariance function (kernel function) is one of the aims of this project.
Boosting the Tied Factor Analysis for Face Recognition Across Large Pose Variations
One of the state-of-the-art methods to make face recognition robust against the change in pose is Tied Factor Analysis. The analysis is applied to each pose to find the linear transformation that relates that pose to the upright pose. This method is a generative statistical approach. In this project we want to use the discriminative Boosting approach to improve the basic Tied Factor Analysis as a weak classifier. The proposed method is evaluated using the standard face recognition dataset such as PIE.