14 / 24
Their work focuses on video
understanding.
Together,
they
published a paper on discovering
objects in videos in an unsupervised
fashion. They have formulated a
solution for efficient and fast object
discovery with a much different
approach than many of the other
current papers. Their paper proposes
a general approach that could work
with any kind of classifier and will
benefit from current neural nets
techniques.
It explores object
discovery using a simple solution
based on common sense and the basic
difference between a foreground and
background object.
They took a completely different
approach to unsupervised learning
which no one else has done before.
They came up with a solution for
learning from highly probable positive
features. This differs from current
approaches that try to find the
differences between the object and
the background. Instead, they took a
new approach that tries to learn what
the object looks like. By knowing the
object well
enough, they can
differentiate the object from the rest
of the background.
When asked about the challenges of
the work, Marius revealed, “I think
that this work is showing a very
interesting secret that we have
discovered regarding unsupervised
learning.” In their view, unsupervised
learning can be done in a video
because it has spatial and temporal
consistencies.
Marius and Ema
14
Friday
Marius Leordeanu is Associate Professor at the University
Politehnica of Bucharest as well as a senior researcher at the
Institute of Mathematics of the Romanian Academy. He supervises
Emanuela Haller, who is currently a PhD student at the University
Politehnica of Bucharest. Marius and Emanuela will present their
poster today (Friday) at ICCV2017.
Unsupervised Object Segmentation in Video by
Efficient Selection of Highly Probable Positive Features