CVPR Daily - Tuesday

8 / 30

the students that are braving the new
world in these days.
Now, in terms of skills. I believe that
this generation is able to get
somewhere faster. Nowadays, they
have more tools, and they can
recombine software pieces. So in a
way it’s exciting, the pace at which
they can go. Perhaps if I can dare to
make a recommendation, something
that I felt that back in the days were
perhaps a little bit better. The students
today have a tendency to be very
rushed to say: I am working on
whatever is hot now, and what
happened three years ago is forgotten.
And this is very short-term sometimes.
So perhaps back in the days, the
students were trying to think a little bit
more like: how can I change things
globally? And they looked a little bit
more beyond their field. So perhaps
this has changed, but this is also a
reaction to the environment. Today
things just have to go quicker.
What is the biggest surprise that you
ever experienced from a student?
Oh, voilà! I would need a lot of time to
answer, because there were so many
times my students surprised me. So
many times! Sometimes positive, and
sometimes negative. But they often
surprise you. And it’s very important -
back to information theory - anytime a
student surprises you, positively or
negatively, and you think - how can
they be so silly, or what a genius - both
times, take a step back as an advisor
and update your own neural network
in your head [laughs ]. Update the
student model, because that’s where
you learn, the surprise. So, I will just
answer with something that comes to
my mind which is fun, it’s part of our
papers we have at CVPR.
It’s a technical contribution, but I
thought it was really fun. My student
was working on this project, where we
try to learn object class detectors by
annotating objects using the center
point, instead of drawing a box around
it. And my student was saying: you
know, we should ask two people to
click in the middle. And I said, forget it!
It’s useless! It’s just a little bit of noise
cancellation. The student said: well,
you know what, Vitto. If they are both
asked to click in the middle, they are
going to make an error. And I said, so
what? So the student said: but the
errors they make is related to how big
the object is. Because if the object is
big, the two annotators are going to
click further apart from each other,
and on the smaller object they will
click closer. And therefore we can
estimate how big the object is based
on the errors the annotators make.
And I was like: that’s awesome! How
did you think of this? We are exploiting
errors to get information about the
object scale out. And in weakly
supervised learning, object scale is one
of the big holy grails. If you have it, it
makes it a lot easier to learn. And so,
you know, we dumped it into this
paper and it became one of the coolest
bits of the paper. I thought - how did
you think of exploiting the errors in
humans? When I think about it, I want
to cancel out the errors, not turn them
into information. I was really
impressed by the student when he said
that. It was very clever: the noise is
noise with respect to the center point,
but it’s signal with respect to scale.
8 Vittorio Ferrari
“How can they be so silly,
or what a genius…”
“That’s awesome!
How did you think of this? ”