CVPR Daily - Tuesday

28 / 30

Sergi Caelles and Laura Leal-
Taixé are both authors of this
CVPR paper, which was joint
Daniel Cremers and Luc Van
Gool. Sergi is a PhD student at
ETH Zürich and Laura is a
postdoctoral researcher at the
Technical University of Munich .
Their work focuses on semi-supervised
video object segmentation: given a
video, the goal is to segment a specific
object for the whole video, given the
first frame. They are the first to
approach this task using deep learning.
When we asked why nobody has taken
this kind of approach before, Laura
says: “Ideas are not so easy to come
by”. However they method itself is not
very complex, she explained, but it’s a
fast method and it clearly outperforms
the state of the art. “Sometimes,
simplicity works really well”, Sergi says.
He told us that one of the most
challenging parts of this work was to
use all the parts of the model in the
right way, although the model
architecture was not very complex.
Their approach consists of a separation
into first training the parent network
with the DAVIS training set, using a
fully convolutional neural network, to
separate objects from the background.
Then they fine-tune on the first
segmentation mask of the video. “This
really is the key point: that you learn
the appearance of the object during
this fine-tuning”, Laura told us.
In the current work, they are
considering the appearance model of
only a specific object. In a follow-up
paper, Sergi tells us, they are
introducing the concept of instance
and segmentation into the mix. They
then not only learn the appearance of
the object, but also the category of the
object and that it is a particular
instance of the object.
Laura Leal-Taixé
Sergi Caelles 28
One-Shot Video Object Segmentation
The authors Laura Leal-Taixé, Sergi
Caelles, Kevis-Kokitsi Maninis and
Jordi Pont-Tuset catching up with our
editor at their poster, yesterday at
CVPR.