Philip H. S. Torr

Professor

Philip H. S. Torr

: http://cms.brookes.ac.uk/staff/PhilipTorr/

 

Microsoft Research: http://research.microsoft.com/virtuamsr/

publications: http://www.informatik.uni-trier.de/~ley/db/indices/a-tree/t/Torr:Philip_H=_S=.html

Key Words

Computer Vision, Virtual Reality, Vanguard, Model Selection, Robust Statistics, Detection of Degeneracy, Motion Segmentation/Clustering.

 

Research Statement

At present I am working on the topics listed below, I envisage that some of these will form the basis for new grant applications, should I move to an academic institution. My goal in research is to choose problems that require strong mathematical theory to yield results that will be commercially useful:

New view synthesis, with emphasis on gaze correction for PC to PC interaction (i.e. developing technologies to facilitate the use of the PC as a video telephone). To this end I have developed new Bayesian stereo algorithms based on the Forward-Backward algorithm used in speech to help reduce the artefacts often present in the synthesized Cyclopean view.

Tracking, currently I am co-supervising two PhD students who are involved in designing a hand tracker. We have developed a new generic Bayesian method for tracking articulated bodies. Uses for this include computer understanding of sign language, a hands free mouse and driving avatars for film special effects.

Markov Random Fields, I have developed a new way for solving Markov random fields using semi definite programming. Currently I am developing Markov random fields for image segmentation, more specifically applied to image and video editing.

Stereo and motion analysis, as well as image geometry. I am working with one PhD student to allow automatic 3D reconstruction of objects from video for people to display as VRML (i.e. 3D reconstruction of archaeological objects etc).I have also worked on Object Recognition, including developing a new algorithm for face detection, to facilitate the organization of digital home photos. One of my interests in how object-recognition can be combined with stereo reconstruction. My PhD student Anthony Dick’s thesis was about the use of domain specific knowledge for creating virtual reality from images of buildings http://svr-www.eng.cam.ac.uk/~ard28/demo2.html. Allowing for automatic reconstruction of architecture from video.

I have just started working on Object Recognition problems; i.e. the problem of discerning what images represent which is perhaps the main problem of computer vision. At Oxford Brookes I have one student Pawan Kumar who is working on this.