I recently obtained my Ph.D. degree in Language and Information Technologies in the School of Computer Science at the Carnegie Mellon University. My Ph.D. advisor was Jie Yang. After graduation, I briefly worked with Howard Wactlar as a Post Doctoral Fellow in the Computer Science Department in the same school at Carnegie Mellon. I also worked with Alex Hauptmann. Before coming to Pittsburgh, I studied in the Tsinghua University (Beijing, China) for my B.Eng. in Computer Science & Technology (2001) and in the National University of Singapore for my M.Sc. in Computer Science (2003) advised by Chin-Hui Lee and Tat-Seng Chua. My current and past research projects focus on multimedia, computer vision and machine learning and their applications. In my spare time, I like spending time with my family, playing tennis with friends, watching movies, reading books, jogging, walking and swimming.
My research interests include multimedia, computer vision, machine learning and their applications. Curriculum Vitae: [PDF];
Resume: [PDF]
PhD Thesis: Multimedia Technologies for Landmark-Based Vehicle Navigation, Carnegie Mellon University, November 2009. [Abstract]
[PDF]
My research interests have been in the intersection between computational theory and practical engineering implementation. I believe that sound mathematical foundations enable engineering successes while practical aspects of real-world implementations inspire theoretical advancement. My research experiences centre around two themes: real-world applications and high performance systems. The themes have driven the development of AI, Human-Computer Interaction and other CS areas for decades. I have researched a number of problems in machine learning systems, multimedia applications, human-computer interface and image processing and computer vision. My research domains range from conventional media (text, image, video, audio) to new media (social network, full-windshield display).
My PhD research endeavors have mainly focused on several intelligent systems for landmark-based vehicle navigation. My research goal is to develop camera-based technologies for next-generation vehicle navigation, which cover landmark labeling, detection, recognition and human vehicle interface. Landmark labeling is essential for development of landmark detection and recognition systems. I have shown that computers can effectively label thousands of images with a small amount of manpower by generalizing Gaussian random field with harmonic functions [ACM MM'06] and our method can reach pixel-level object extraction performance [IEEE T-IP'09]. Road signs are a special class of landmarks that carry critical guidance information for drivers. Inspired by this observation, I have developed an automated system that detects and recognizes text on road signs from real driving videos [IEEE T-ITS'05, ACM MM'04]. Humans not only read road signs but also utilize other landmarks (e.g., traffic lights, bridges, store signs and/or buildings) to help themselves navigate when walking and driving. However, most current GPS navigation systems rarely provide capabilities beyond the limitations of a map dataset and satellite signals to help drivers navigate. We have developed systems that can automatically recognize store signs and landmark buildings from driving videos using our proposed object fingerprint framework [Submitted, ACM MM'08]. We have demonstrated above research outcomes on computer displays [ICME'06] and on a state-of-the-art full-windshield display system [ACM MM'09]. The algorithms and technologies developed in my thesis can be applied to other human centered computing applications.
In summer 2008, I participated in a collaborative endeavor with Intel Research, Pittsburgh - building an image dataset of 101 fast foods from over 10 food chains. I did this summer project and research as a V-Unit. The built PFID dataset contains 30,603 still images, 606 stereo pairs, 303 360-degree food videos and 27 privacy-preserving videos of eating events [ICIP'09]. Self-reporting, the main method for dietary data acquisition, underestimates food intake and do not reflect the real dietary info. This observation inspired us to build PFID and develop fast food recognition methods [ICME'09]. Through this project I have deepened my understanding of how society grows science and technology, and how in turn science and technology changes society and the world. My PhD research has been supported by General Motors and I have also briefly worked on a few projects sponsored by NSF, NIH, Microsoft, etc.
The symbiosis of computational theory and engineering has already had a long-standing impact on a classic area - text mining. My research has shown that, with right models and representations, valuable knowledge can be successfully discovered from various text forms [CEAS'07, ACM TOIS'06, ICML'04, SIGIR'03]. Take a recent work as an example. Using only traffic-based patterns extracted from email headers (e.g., frequency counts, etc), we have applied standard data mining and classification algorithms to obtain 94% accuracy in predicting the leadership roles of group members in the CMU GSIA corpus containing 15K emails [CEAS'07]. This symbiosis, which is vital to tackle many fundamental and interesting problems in the areas of intelligent systems, human-computer interaction, data mining, etc, has also led to promising outcomes in my other endeavors [CIVR'08, ICME'05].
Some Teaching Experiences
Teaching Assistant, Course: 15-384 Robot Manipulation
Aug 2006 - Dec 2006, School of Computer Science, Carnegie Mellon University
Jobs (co-work with the instructor and the other TA): office hours, grading homework and exams, designing tournament projects, organizing demos.