3D Object recognition with local features
Date: 11:30am, 8th Mary, 2006
Location: ITL Meeting room top floor
Speaker: Dr Francesca Odone, University of Verona, Italy

Many classification tasks based on visual cues can be
successfully addressed by first extracting meaningful information
from images, then finding descriptions based on this information,
and finally designing classification algorithms able to discriminate
between the classes of interest. The image descriptions should
maximize the interclass distance and minimize the intraclass
variability. On this respect, in the last years a huge amount of
work on finding image keypoints robust to environment and viewpoint
variations have been carried out.

In this talk I will address 3D object recognition, proposing a
method based on image description with scale invariant local
keypoints, and recognition with a collection of Kernel-based
classifiers.

One of the main challenges of this approach is due to the
variable-length descriptions obtained from local keypoints.  I will
describe a "bag of keypoints" approach to this problem, reporting
promising recognition results. I will also discuss the connections
between this approach and very recent works on kernel engineering
for local features, highlighting the pros and cons of the two
choices.

(Joint work with E. Arnaud, E. Delponte, A. Verri)