Paolo Piro
Currently I am PostDoc researcher at the Italian Institute of Technology (IIT),
in the Computer Imaging facility.
My PostDoc research is in computer vision and machine learning, and is mainly focused on the following topics:
Recently I have defended my Ph.D. thesis at the University of Nice-Sophia Antipolis,
I3S Laboratory, under the supervision of
Prof. Michel Barlaud.
My Ph.D. thesis was also in co-tutorship with
Prof. Giulio Iannello
of the Campus Bio-Medico University of Rome.
I have collaborated with Frank Nielsen
(LIX, Ecole Polytechnique), Richard Nock
(CEREGMIA, University of Antilles-Guyane), Eric Debreuve (I3S/CNRS, University of Nice-Sophia Antipolis) and
Sandrine Anthoine (LATP, University of Provence/CNRS).
My Ph.D. thesis is mainly focused on the following topics:
Ph.D. thesis
|
Paolo Piro.
"Learning prototype-based classification rules in a boosting framework:
application to real-world and medical image categorization". University of Nice-Sophia Antipolis, January 2011. |
|

Boosting k-NN
We have generalized the classic k-NN classification in a supervised learning framework.
Namely, we have proposed an instance-based classification rule, which can be induced
by our simple UNN algorithm minimizing a convex risk function on training data.
Our learning algorithm is inspired by AdaBoost in that it linearly combines predictions
from learned prototypes.
UNN enables consistent data reduction and is adapted to multi-class classification.

Medical image retrieval
We have addressed the problem of annotating medical images for
computer-aided diagnosis.
In particular, we have proposed a novel multi-class leaning algorithm, MLNN,
for leveraging prototype-based classification.
We have tested MLNN for automatic body part recognition on a set of medical radiographs
taken from the clinical routine. Automatic indexing of such images is expected to be
the first step towards a full computer-aided diagnosis system.
ICOS-HD project
My PhD thesis has been part of the ANR research project Icos-HD.
The main goal of Icos-HD was to define a suitable framework for scalable content-based description of HD images and videos.
In particular, we concentrated on low-level visual image descriptors, and proposed a new descriptor, called SMP.
SMP exploits space and scale correlations of Laplacian pyramid coefficients for representing visual information
in a sparse way.
Bregman nearest neighbors
Nearest neighbor (NN) search is a crucial tool in content-based image indexing and retrieval.
We have addressed the problem of searching NN efficiently when classic distance metrics are replaced by
Bregman divergences.
Namely, we have considered two data structures that have proven very effective in applications.
First, we have improved Bregman Ball trees, also adapting them to symmetrized divergences.
Then, we have generalized vantage point trees to Bregman divergences.





