Probabilistic Models for Sequences

This project was funded by EPSRC under grant GR/L78161, with additional support from BAE Systems. The project duration was October 1998-September 2001.

People involved with the project were Amos Storkey, Stephen Felderhof and Chris Williams.

Abstract

The major achievement of this project was to develop the Position Encoding Dynamic Tree (PEDT) hierarchical model for the analysis of single images and image sequences. This architecture is designed to produce a ``parse tree'' type interpretation of images, capturing the concepts of objects, spatial locality, component/subcomponent relationships and the propagation of this structure through time. Given an image sequence, inference in this model is NP-hard, but we have developed a structured variational approximation method for inference, and shown that it gives superior performance over the mean field variational method. The model has been applied to a number of image labelling tasks on both single images and image sequences (including a road-scene analysis task) and has demonstrated improved performance.

In addition, research under the project also compared mixture models and neural networks as pixel models, investigated products of tree-structured Gaussian experts as image models and developed a hierarchical Gaussian density model (giving a well-founded probabilistic hierarchical clustering model).

Publications

2003

Dynamic Trees for Image Modelling pdf
Nicholas J. Adams, Christopher K. I. Williams
Final draft of paper appearing in Image and Vision Computing 20(10) 865-877 (2003)

2002

Image Modelling with Position-Encoding Dynamic Trees pdf
Amos J. Storkey, Christopher K. I. Williams
Final draft of paper appearing in IEEE Pattern Analysis and Machine Intelligence 25(7) 859-871 (2003).

Dynamic Trees: Learning to Model Outdoor Scenes gzipped postscript
Nicholas J. Adams, Christopher K. I. Williams
To appear in: Proceedings of the European Conference on Computer Vision 2002. Lecture Notes in Computer Science. © Springer-Verlag (2002)

Products of Gaussians gzipped postscript
Christopher K. I. Williams, Felix V. Agakov, Stephen N. Felderof
To appear in: Advances in Neural Information Processing Systems 14 eds. T. G. Diettrich, S. Becker, Z. Ghahramani MIT Press (2002)

2001

Position Encoding Dynamic Trees for Image Sequence Analysis gzipped postscript
Stephen N. Felderhof, Amos J. Storkey and Christopher K. I. Williams
Technical Report, December 2001, revised version January 2002.
Link to tech report web page with mpeg videos, etc.

Comparing Mean Field and Exact EM in Tree Structured Belief Networks gzipped postscript
Nicholas J. Adams, Christopher K. I. Williams and Amos J. Storkey
In Proceedings of Fourth International ICSC Symposium on Soft Computing and Intelligent Systems for Industry, June 2001.

Products and Sums of Tree-Structured Gaussian Processes gzipped postscript
Christopher K. I. Williams and Stephen N. Felderhof
In Proceedings of Fourth International ICSC Symposium on Soft Computing, June 2001.

Combining belief networks and neural networks for scene segmentation gzipped postscript
Xiaojuan Feng, C. K. I. Williams and S. N. Felderhof
Submitted to IEEE Trans PAMI, April 1999. Revised version March 2001, accepted for publication July 2001.
IEEE legal stuff This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without explicit permission of the copyright holder.

Dynamic Positional Trees for Structural Image Analysis gzipped postscript
Amos J. Storkey and Christopher K. I. Williams
In: Proceedings of the Eighth International Workshop on Artificial Intelligence and Statistics (2001)

2000

MFDTs: Mean Field Dynamic Trees gzipped postscript
Nicholas J. Adams, Amos J. Storkey, Zoubin Ghahramani and Christopher K. I. Williams
In: Proceedings of 15th International Conference on Pattern Recognition, 2000.

Dynamic Trees: A structured variational approach giving efficient propagation rules postscript
A. J. Storkey
in Proceedings of the Sixteeth Conference on Uncertainty in Artificial Intelligence (UAI 2000). Eds. C. Boutilier, M. Goldszmidt, Morgan Kaufmann (2000).
A MCMC approach to Hierarchical Mixture Modelling gzipped postscript
C. K. I. Williams
In Advances in Neural Information Processing Systems 12, eds. S. A. Solla, T. K. Leen and K-R. Muller, MIT Press (2000)

Software

The PEDT sofware produced in the project is available at http://www.anc.ed.ac.uk/code/storkey/.

Software for (non position encoding) Dynamic Trees written by Nick Adams as part of his EPSRC-funded PhD is available at http://www.anc.ed.ac.uk/code/adams/dt/dt.tgz

This page is maintained by Chris Williams (ckiw@dai.ed.ac.uk)