Deformable Objects

When?

Wednesday 10 October 2012, 14:00 to 15:00

Where?

CVSSP Seminar Room (40bAB05)

Open to:

Public, Staff, Students

Speaker:

Dr. Xianghua Xie, University of Swansea

In this seminar, I will first talk about deformable modelling for object segmentation. Deformable models have been widely used for shape recovery due to their natural handling of shape variation and the ability to incorporate prior knowledge. Their design normally involves the consideration of the following three fundamental issues: representation and its numerical solution, object boundary description and stopping function design, and initialisation and convergence. These issues are not always independent of each other. An appropriate representation ensures the model handle the deformations properly, however, representation schemes that support topological changes do not necessarily always achieve the desired evolutions. Better boundary description can improve convergence ability by preventing leakage through and good convergence ability is paramount to not compromise any gains from carefully chosen representation and boundary extraction. On the other hand, good convergence properties can compensate certain inadequacies in boundary description. In this talk, I would like to show some of our recent developments in implicit deformable modelling in 2D, 3D, 3D+time and object tracking, and discuss their relationship to above three design issues. The talk will cover our novel implicit representation which can handle more complex topological changes than splitting and merging, an edge based physics-inspired external force field and its generalisation to arbitrary dimensions that has superior performance in capturing complex geometries and dealing with weak edges and broken boundaries, a simple but effective numerical solution to avoid periodic regularisation so that initialisation is no longer necessary which is particularly useful for object detection, a 4D segmentation in SPECT which incorporates learnt shape and appearance prior, and an application to object tracking.

I will also briefly present our recent work on 3D human pose estimation by exploiting the advantages of fixing the root node. This approach has two key benefits: The first is that each local solution can be found by modelling the articulated object as a kinematic chain, which has far less degrees of freedom than alternative models. The second is that by using this approach we can represent, or support, a much larger area of the posterior than is currently possible. This allows far more robust algorithms to be implemented since there is far less pressure to prune the search space to free up computational resources.

Speaker Biography:

Dr. Xianghua Xie is currently a senior lecturer in the Department of Computer Science, University of Swansea. He received the M.Sc. (with commendation) and Ph.D. degrees in Computer Science in 2002 and 2006 respectively, from the University of Bristol. In 2007, he took an RCUK Academic Fellowship at Swansea University. His current research interests are video analysis, human pose estimation and tracking, human interaction modelling, texture analysis, image segmentation, surface inspection, deformable models, and level set methods.