
180 Image Processing for Cinema
that there usually is a wide range of possible smoothed camera motions that
the viewers may accept as providing a stable enough result. From these ob-
servations, the authors introduce an algorithm that starts by extracting a set
of robust feature trajectories and then finds a set of possible 2D transforma-
tions that smooth them; from these possible transformations, they choose the
one that avoids leaving large blank areas after image composition. Features
are found with SIFT [258], linked with neighbors using a Delaunay triangu-
lation [106], and propagated by a feature matching procedure that also takes
into account ...