Research

Robust Face Recognition

Face recognition research is into its fourth decade now, yet the problem remains unsolved in general.  To be sure, we have face recognition systems deployed at airports and city centers, and they work reasonably well.  The best systems today are still not sufficiently robust to handle simultaneous changes in illumination, head pose, and facial expression, to name a few. 

We propose a synthesis approach to achieve robustness: from a few training images of an individual, we synthesize more images of the same person under all kinds of variations: illumination, pose, expression, etc.  From this enlarged set of training images, we explore classifiers that can effectively learn a person's identity in a robust manner.  Our approach combines graphics techniques (image-based rendering) and pattern recognition ideas.  More details here.

Multimodal Biometrics

Biometrics (e.g. fingerprint, iris, face patterns) is increasingly being used to enhance security.  Recent high profile projects include the biometric passport project by the Singapore Immigration and Checkpoints Authority.  Each type (modality) of biometrics has its pros and cons, and much ongoing research is being pursued improve its performance.  Multimodal biometrics refers to the combination (fusion) of different modalities.  This not only improves accuracy, but also makes the overall system more difficult to thwart.  We have developed a prototype system that combines face and fingerprint biometrics to continuously verify that the user in front of a computer is who he claims to be.  The system can take appropriate action if the user is absent, or if an impostor has taken over the computer (hijacked).  More details here.

Computational Photography

Digital photography is very popular now.  Yet there are obvious problems: red-eye, uneven exposure, blur, etc.  We seek to understand and correct these problems using a combination of image-based rendering techniques and devising new hardware.  More details here.

Music Transcription

Music transcription refers to the "conversion" of music from an acoustic signal into symbolic form, for example, going from a .wav file to a MIDI file.  This is a typical "inverse problem" and it is hard!  We propose using an instrument model combined with statistical techniques to transcribe polyphonic and multi-instrument music.

Other projects

Smile!  Wear it like a costume.

Staff and Students

• Guo Dong, Ph.D student. Face recognition, Computational photography.
• Zhuo Shaojie, Ph.D student. Computational photography.
• Ye Ning, Ph.D student. Face modeling and animation.
• Ha Mai Lan, Ph.D student. Computational photography.
• Hossein Nejati, Ph.D. student. Face recognition.
• Vlad Hosu, Ph.D. student. Computational photography.
• Lewis Foo, Ph.D. student. Computational photography.
• Ma Keng Teck, Ph.D. student. Scene understanding.
• Hamed Kiani Galoogahi, Ph.D. student. Biometrics.

Previous Staff and Students

• Dr. Chen Yan, Post-doctorate Fellow. Extreme face recognition.
• Dr. Yu Dan, Post-doctorate Fellow.  Extreme face recognition.
• Rajkumar Janakiraman, Research Assistant.  Face recognition, Continuous Authentication.
• Feng Jimin, Research Assistant.  Extreme face recognition.
• Zhang Sheng, Ph.D. student.  Face space, face recognition.
• Miao Xiao Ping, Ph.D. student.  Computational photography.
• Yin Jun, Msc student.  Music transcription.
• Huang Xiao Yu, Msc student.  Head detection and tracking in crowds.
• Zhang Xiao Peng, Ph.D. student. Computational photography.

Last updated September 2010.