CVPR Daily - Tuesday

18 / 30

Nour Karessli is a computer vision
engineer at EyeEm, who are located in
Berlin. She published “ Gaze
Embeddings for Zero-Shot Image
Classification” here at CVPR, together
with , Bernt Schiele and
Andreas Bulling.
Nour, where are you working at the
moment?
I work at EyeEm, which is a
photography company, and we work on
cutting-edge technology for computer
vision. We connect a community of
talented photographers with iconic
brands and sell photos. I finished my
master’s degree in July last year and
started at EyeEm in August, so it’s been
a year.
What was the focus of your master?
My master thesis was about gaze
embeddings for zero-shot learning for
classification. I did it at the Max Planck
Institute in Saarland.
I understand you did not start your
studies there?
I started my master’s there, and before
that I was doing a bachelor in Syria, at
the Damascus University.
You are doing a presentation today.
What is the work that you are
presenting?
The work I am presenting is a paper
about gaze embedding for zero-shot
image classification. In this paper we
use human gaze information to guide
the classification task in a zero-shot
setting. We make use of the human
ability to distinguish between different
classes unconsciously.
What is the novelty of this work?
Previous approaches used object
discriminative properties collected by
experts, and then the annotators had to
go through the objects and annotate
these attributes. This is very costly,
especially for fine-grained classification.
It’s also difficult because the categories
are visually very similar, and thus our
suggestion is to use the gaze
information. It's cheaper and faster,
because it's implicit. You just ask the
annotator to look at the image and
distinguish between the objects, and
then the human - without thinking
about it - will focus on the important
features.
Women in Computer Vision
18 Nour Karessli
“We make use of the human
ability to distinguish between
different classes unconsciously”