Research Summary

Changjiang Yang
Computer Vision Laboratory, Department of Computer Science
University of Maryland, College Park, MD 20742, USA

My research has spanned three areas that are closely related to each other: computer vision, machine learning and scientific computing. The goal of computer vision is to build a machine which can see the world as human beings. To achieve this goal, some form of human intelligence is indispensable, just as people see the outside world by understanding it first. On the other hand, computer vision is a very important research area for the machine learning, and it has a large amount of applications. Usually the real-world problems in vision and learning are complicated, high-dimensional, non-linear and large-scale. To solve these problems, efficient algorithms are required. Therefore scientific computing has played an important role in vision and learning, and it will continuously be more important. My research contributions can be divided into the following closely related topics.

• Fast Gauss transform has been studied and we find that it is inefficient in higher dimensional spaces. We developed a new improved fast Gauss transform using algorithms from approximation algorithms and numerical analysis. The improved fast Gauss transform has been applied to various applications, such as efficient kernel density estimation, image segmentation, image registration, object tracking, etc.
• Many methods in vision or learning require the evaluation of the similarity measures between two probability distributions. The commonly used similarity measures such as Kullback-Leibler distance and Bhattacharyya distance are limited to one or two feature dimensions, due to the difficulty in estimating the entropy of the high-dimensional features. We proposed a similarity measure which is the sum of all pair-wise kernelized distances between two distributions. It can be efficiently computed using our improved fast Gauss transform.
• Visual tracking is an important and active research area in computer vision. We used our similarity measure and mean-shift technique to track single object in a joint feature-spatial space, and achieved more accurate and robust tracking performance. For multiple object tracking, we used distinctive features and particle filter framework to simultaneously and reliably track them. The quasi-random sampling and efficient distinctive features are used to achieve a realtime tracking speed.

We have submitted and published these work in computer science and applied mathematics journals and conferences. It is also available as technical reports in Department of Computer Science, University of Maryland at College Park.

The C++ code of improved fast Gauss transform is available for any non-commercial,  academic purpose without fee. Please read the license first if you want to download the code (Click here and there is a link at the end of the license). Please send email to yangcj@cs.umd.edu, ramani@umiacs.umd.edu and lsd@umiacs.umd.edu if you have any question.