Ph.D. Research Proposal Exam: Jaishanker K. Pillai
Friday, April 27, 2012
2:00 p.m. Room 3450, AV Williams Bldg.
For More Information:
301 405 3681 firstname.lastname@example.org
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Jaishanker K. Pillai
Professor Rama Chellappa
Professor Larry Davis
Professor Min Wu
Date/Time: Friday April 27th, 2012, at 2pm
Location: Room 3450, AV Williams Building
Title: Learning Visual Classifiers using Limited Labeled Images
Abstract: Recognizing humans and their activities from images and video is one of the key goals of computer vision. Supervised learning algorithms like Support Vector Machines and Boosting often require large amount of labeled data for good performance. While labeling visual data is difficult due to the significant human effort involved in annotation, it is often considerably easier to collect unlabeled data due to the availability of inexpensive cameras and large public databases like Flickr and Youtube. In this proposal, we develop efficient machine learning techniques for visual classification from small amount of labeled training data by utilizing the structure in the testing data, unlabeled data or labeled data in a different domain.
In the first part of the proposal, we consider how multiple noisy samples available during testing can be utilized to perform accurate visual classification. Specifically, we study the problem of unconstrained human recognition from iris images. We develop a Sparse Representation-based selection and recognition scheme, which learns the underlying structure of clean images and propose a quality-based fusion scheme to combine the varying evidence.
In the second part of the proposal, we demonstrate how labeled data in a different domain can aid visual classification. We consider the problem of shifts in acquisition conditions during training and testing, which is very common in iris biometrics. We provide one of the first solutions to this problem, a kernel learning framework to adapt iris data collected from one sensor to another. The proposed method produces significant improvement in accuracy, is robust to occlusions, rotations and can incorporate privacy using cancelable iris patterns.
In the third part, we analyze how unlabeled data available during training can assist visual classification applications. Here we consider still image-based vision applications involving humans and extract the implicit motion information in human poses using unlabeled videos. We pose this inference of implicit motion information as a non parametric density estimation problem on non-Euclidean manifolds. Our experiments illustrate that the extracted motion information benefits a variety of applications in computer vision like activity recognition, video summarization and motion prediction.