Supplementary Material for:
On Compositional Image Alignment
with an Application to Active Appearance Models

Contents

This is the supporting online material for the CVPR 2009 paper on compositional image alignment. The most up to date version of the data with high quality movies can be found online. It contains

Talking Face Fitting

The video shows the results of tracking a 5000 frame sequence with the compositional methods

Tracking a talking face

Our algorithm makes fast and robust tracking possible. We compare face tracking under natural motion, using ICIA, LinCoDe and CoDe. The original ICIA fails immediately with this large model and new face data. Substituting the orthonormal incremental warp for the original ICIA warp, the algorithm still loses track very early, whereas LinCoDe and CoDe can track much further. Finally, adding regularisation to all algorithms, ICIA still loses track completely after approximately 500 frames and does not recover the local deformations accurately. In contrast CoDe now tracks the full 5000 frame sequence without reinitialization, and LinCoDe tracks for 2500 frames.

The video is encoded with the msmpeg4v2, which should be installed on linux and windows, and can be downloaded from mediacodec.org.

Low Resolution Talking Face Fitting

The video shows the results of tracking a low resolution face with lots of out of plane rotation and strong expressions with the compositional methods

Tracking a low resolution talking face

Tracking a low resolution video with large head motions succeeds with CoDe, where ICIA fails. All methods used the orthonormal incremental warp, and relatively strong regularisation. ICIA starts to drift in the early frames, while CoDe tracks the full sequence. The approximate gradient method LinCoDe also suceeds, but looses track of the details for about 100 frames.

The video is encoded with the msmpeg4v2, which should be installed on linux and windows, and can be downloaded from mediacodec.org.