Previous Ph.D ProjectsThe following is a selection of previous Ph.D projects I have supervisedJonathan Teh - Object-based Video Compression A novel technique for
region based motion compensation has been developed based on
applying a robust motion compensation technque using an overlapping
block-based tessellation of each frame and representing region
boundaries with a piece-wise B-spline model. This motion compensation
technique has been integrated into a wavelet-based video compression
system.
David Gibson - Motion Trajectory Estimation in Long Image Sequences. This research is
aimed at estimating a set
of motion trajectories over a large number of frames in an image
sequence.
It could be regarded as being in between traditional dense optical flow
estimation and feature point tracking, the latter producing a set of
sparse
displacement estimates typically at object edge and corner points. This
work, uses the
distinctiveness of spatial texture patterns and contextual contraints
of
the parameterised trajectories (integrated into a Markov Random Field
formulation)
in order to robustly compute the trajectories which are not limited
simply
opf object boundaries [Gibson-Spann
1999].
Applications of this work are in motion
compensated
predication for video compression and optical flow estimation.
Klaus Koester - A Robust
Statistical Approach to Image Segmentation
In this research we have applied robust regression estimation techniques to the problems of range image segmentation and to the segmentation of 3D seismic data. The segmentation algorithm applies an overlapping block strategy which adapts to the local structure of the image thus effectively increasing the % breakdown ratio of the robust estimator [Koester-Spann 2000]. A statistic known as the Mutual Inlier Ratio (MIR) is used to control the local pixel/voxel clustering. This statistic is compared favourably to the traditional KS statistic for comparing distributions. For the case of the range image segmentation, an extensive experimental comparison has been carried out with a number of previously published methods using a software tool to compute segmentation errors and results show that our method compares favourably with all other methods used in the evaluation and being particularly effective on noisy data. E. P. Ong - Robust Optical
Flow Estimation
The problem of motion estimation is one of estimating a smoothly varying optical flow field whilst being able to take into account both structure and motion boundaries leading to disconitnuities in the optical flow. We have successfully solved this problem using robust regression techniques applied to computing flow over an overlapping block tessellation [Ong-Spann 1998]. A thorough experimental evaluation of this work has been carried out to compare it with previously published techniques and we have found that, amongst optical flow estimation algorithms that result in a dense flow field (defined as one resulting in more than 80% spatial coverage) our algorithm outperforms all other methods tested irrespective of the error metric used. |