Ph.D. Candidate

Robotics Institute

School of Computer Science

Carnegie Mellon University

Pittsburgh, PA 15213

My Carnegie Mellon homepage

I am now a Ph.D. student at Carnegie Mellon University, the Robotics Institute in the School of Computer Science.  I am co-advised by Professor Takeo Kanade and Martial Hebert.  Before coming to Pittsburgh, I got my Master and Bachelor degrees both from the University of Science and Technology of China (USTC).

My Research Interests include: Computer Vision, Machine Learning, HCI, Multimedia Processing and Data Mining.  I am generally interested in things cool, things that can change what our everyday life looks like.

09/2003 ~ 11/2004,  Research Assistant at USTC

12/2004~02/2006, Visiting student at MSRA

Video Frame Representativeness Learning (05/2005~07/2005)
Hong-Wen Kang, Xian-Sheng Hua. "To learn Representativeness of Video Frames", In proceeding of ACM International Conference on Multimedia (ACM MM), Singapore, Nov. 2005.  [PDF 778KB]

Space-Time Video Summarization (07/2005~11/2005)
Hong-Wen Kang, Yasuyuki Matsushita, Xiaoou Tang, Xue-Quan Chen. "Space-Time Video Montage", In proceedings of  IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) 2006.[PDF 1.54MB]

Some examples: Space, Time, Space_Time, Multiple Input, Failure Case

Machine Learning by Carlos Guestrin, Spring 2007

Learning-based Methods in Vision by Alyosha Efros, Spring 2007

Mathematical Fundamentals for Robotics , by Michael Erdmann, Fall 2006

Final project: GPU Based Space-Time Video Packing.  Video demo [WMV 10.4MB]

In this project, I looked at using Graphics Processing Unit (GPU) for simultaneous 3D objects packing.  As an application, I used the method in the former work of Space-Time Video Montage.  Comparing with original sequential-greedy packing, the new method makes better use of the output volume and generates more densely and evenly packed results.   At the same time, according to initial experiments, there is no increase in computational cost. 

The following figure compares the result from sequential First-Fit method and the new GPU based method. a) and b) are two input video frames, c) shows the result of sequential First-Fit method, d) shows the result of the new simultaneous packing method.

Computer Vision, by Martial Hebert, Fall 2006

Final project: Video Google: a video retrieval system. In collaboration with Daniel Leeds, we implemented a video retrieval system based on the paper: Video Google: A Text Retrieval Approach to Object Matching in Videos by Josef Sivic and Andrew Zisserman.

Some results from our system:
*Retrieved results are the first 9 frames, with descend relevance.
*Video data from: Dressed To Kill (1946, B&W), directed by Roy William Neill.

Test 1,

Query

Retrieved frames and objects.

Test 2,

Query

Retrieved frames and objects.