SUPPLEMENTARY MATERIAL FOR VMV ARTICLE:
Direct Image Registration without Region of Interest
F. Brunet, A. Baroli, N. Navab, R. Malgouyres

There are two videos in the ZIP file provided with our submission:
- phdcomics.avi
- note.avi
These two videos are encoded using the standard XviD codec. They are thus 
easily readable with the most common players such as MS Windows Media Player,
VLC, MPlayer, etc.


*******************************************************************************
phdcomics.avi

This short video is a simple demonstration of our algorithm with a Free-Form
Deformation model. A video sequence was acquired with a standard camera. Each
frame of the video sequence is then registered to the first frame (the
registration algorithm for the frame #i is initialized with the warp obtained
for the frame #(i-1). The overlap automatically determined with our approach is
also shown (the pixels not in the overlap are coloured in red). The "difference
image" is obtained by substracting the "warped target" to the "source image".


*******************************************************************************
note.avi

In this second video, we compare several approaches to image registration:
- our approach
- a fixed rectangular RoI with a large margin (30 pixels), i.e. a small RoI
- a fixed rectangular RoI with a small margin (10 pixels), i.e. a large RoI

The experimental setup is similar to the one used in the "phdcomics.avi" video.

This video illustrates the fact that an approach with a fixed RoI that is not
completely included in the overlap between the images can lead the registration
algorithm to fail. It also shows that our approach copes with all sort of
outliers (`real' outliers such as occlusions and specularities and
`theoretical' outliers: off-target pixels).


*******************************************************************************
The Free-Form Deformation models of the two videos all relies on uniform
bicubic B-splines with 5*5 control points.