Accompanying CD-ROM The CD-ROM inside the back cover of this thesis contains several video sequences referenced in the text. The video sequences are referenced according to their filenames on the CD-ROM:
tracked/V1_graph.m1v
, video V1 along with a plot of the tracked parameters,
tracked/aam_orig.m1v
, video comparing an original video sequence with the resynthesised frames from the tracked parameters,
examples/V1/V1_orig.m1v
, the original video sequence V1,
examples/V1/V1_track.m1v
, video V1 resynthesised from tracked parameters,
examples/V1/V1_arp.m1v
, generated from video V1 using the autoregressive process,
examples/V1/V1_wor.m1v
, generated from video V1 using our model without residuals (greedy algorithm),
examples/V1/V1_wr.m1v
, generated from video V1 using our model and a linear residual model (greedy algorithm),
examples/V1/V1_ncwor.m1v
, generated from video V1 using our model without residuals (normalised cuts algorithm) ,
examples/V1/V1_ncwr.m1v
, generated from video V1 using our model and a linear residual model (normalised cuts algorithm),
examples/V2/V2_orig.m1v
, the original video sequence V2,
examples/V2/V2_track.m1v
, video V2 resynthesised from tracked parameters,
examples/V2/V2_arp.m1v
, generated from video V2 using the autoregressive process,
examples/V2/V2_wor.m1v
, generated from video V2 using our model without residuals (greedy algorithm),
examples/V2/V2_wr.m1v
, generated from video V2 using our model and a linear residual model (greedy algorithm),
examples/V3/V3_orig.m1v
, the original video sequence V3,
examples/V3/V3_track.m1v
, video V3 resynthesised from tracked parameters,
examples/V3/V3_arp.m1v
, generated from video V3 using the autoregressive process,
examples/V3/V3_wor.m1v
, generated from video V3 using our model without residuals (greedy algorithm),
examples/V3/V3_wr.m1v
, generated from video V3 using our model and a linear residual model (greedy algorithm);
exp/m1v/q01.m1v
to exp/m1v/q52.m1v
are the videos used for the psychophysical experiment (in mpeg format);
exp/gif/q01.gif
to exp/gif/q52.gif
are the videos used for the psychophysical experiment (in animated gif format).
index.htm
.
Abstract
Statistical appearance models are used to model objects from images using their shape and texture. Such models have been applied successfully in a large number of applications. Nevertheless, the appearance model does not model video sequences of animated deformable objects.
The aim of this thesis is to add a temporal dimension to the appearance model in order to properly represent movement in video sequences. We apply this extended model to the study of facial behaviour.
The method uses a statistical framework learnt from a training video sequence. The series of parameters extracted from the sequence is modelled by a set of pathlets in parameter space. A higher level model learns how to organise the pathlets into meaningful sequences representing facial expressions or typical movements of the head.
A measure of quality of generated video sequences is derived. This measure shows that our model outperforms an alternative based on autoregressive processes. A forced choice psychophysical experiment confirms this conclusion. Acknowledgements I would like to thank my supervisor, Dr Tim Cootes for his useful advice, his encouragement, guidance and support over the past four years.
I would like to thank all the members of the Imaging Science and Biomedical Engineering group for providing such a good work environment.
I would like to thank my parents for supporting me.
Finally, I would like to thank all the people who volunteered for the psychophysical experiment set up to assess this thesis' framework. I would particularly like to thank Lilian, Nicolas, Fabrice, Juana, Gilles, Vivek, Jun, Alexandre, Arnaud, Bruno, Laurent, Fernand, Marie-Jeanne, Paul, José, Kostas, Domitille, Xavier, Sylvie, David, Kolawole, Roy, John, Mike, Patrick and Panachit for their help and their patience. Publications Some of the work described in this thesis has also appeared in: