Challenges in Scene Understanding
Speaker: Lior Wolf , MIT - CBCL
Date: April 6 2005
Time: 2:45PM to 3:45PM
Location: 32-D507
Host: Greg Shakhnarovich, CSAIL
Contact: Greg Shakhnarovich, xx3-8170, gregory@csail
Relevant URL: *** Note the one-time unusual location: 32-D507
ABSTRACT:
In the past few years, significant progress has been made in the
development of a new technology for supervised learning and in
its use as the core of several real-time vision systems for
detecting specific classes of objects, including people, faces
and cars, within complex images. The time is now ripe to go
beyond detection of single objects and to approach the problem of
scene understanding. This was the original dream of machine
vision: to create a system that could describe what it sees, or
in David Marr's words, what is where. We are currently taking
initial steps in that direction. A system is being built that
should give a coarse description of what appears in an image of a
street after learning from a database of similar images.
In my talk, I will first present a novel set of features for
robust object recognition, which exhibits outstanding performance
on a variety of object categories while being capable of learning
from only few training examples. These features are motivated by
Poggio's quantitative model of visual cortex. The system built on
top of these features outperforms state-of-the-art systems on a
variety of object image data sets.
I will then present and analyze a novel regularization technique
based on enhancing our data set with corrupted copies of the
original data. The motivation is that since the learning
algorithm lacks information about which parts of the data are
reliable, it has to produce more robust classification functions.
Using this framework, I will propose a simple addition to the
gentle boosting algorithm which enables it to work with only few
training examples. I will conclude with several experiments on
vision and biological data sets.
See other events that are part of MIT Machine Vision Colloquium 2004/2005
See other events happening in April 2005
