Deep
neural
networks
achieve
state-of-the-art
performance
on
many
image
segmentation
tasks.
However,
the
nature
of
learned
representations
used
by
these
is
unclear.
Biological
brains
solve
this
task
very
efficiently
and
seemingly
effortlessly.
Neurophysiological
recordings
have
begun
to
elucidate
underlying
mechanisms
segmentation.
In
particular,
it
has
been
proposed
that
border
ownership
selectivity
(BOS)
first
step
in
process
brain.
BOS
a
property
an
orientation
selective
neuron
differentially
respond
object
contour
dependent
location
foreground
(figure).
We
explored
whether
deep
use
close
those
biological
brains,
particular
they
explicitly
represent
BOS.
therefore
developed
suite
in-silico
experiments
test
for
BOS,
similar
probe
primate
tested
two
trained
scene
tasks
(DOC
[1]
Mask
R-CNN
[2]),
as
well
one
network
recognition
(ResNet-50
[3]).
Units
ResNet50
predominantly
showed
contrast
tuning.
responded
weakly
stimuli.
DOC
network,
we
found
units
earlier
layers
stronger
tuning,
while
deeper
increasing
tuning
seems
wide-spread
extrastriate
areas
most
common
intermediate
area
V2
where
prevalence
neurons
exceeds
(V1)
later
(V4)
areas.
also
which
was
natural
images,
did
not
generalize
simple
stimuli
typically
experiments.
This
differs
from
findings
responses
are
than
complex
scenes.
Our
methods
general
can
be
applied
other
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),
Год журнала:
2020,
Номер
unknown, С. 14636 - 14645
Опубликована: Июнь 1, 2020
Current
methods
for
depth
map
prediction
from
monocular
images
tend
to
predict
smooth,
poorly
localized
contours
the
occlusion
boundaries
in
input
image.
This
is
unfortunate
as
are
important
cues
recognize
objects,
and
we
show,
may
lead
a
way
discover
new
objects
scene
reconstruction.
To
improve
predicted
maps,
recent
rely
on
various
forms
of
filtering
or
an
additive
residual
refine
first
estimate.
We
instead
learn
predict,
given
by
some
reconstruction
method,
2D
displacement
field
able
re-sample
pixels
around
into
sharper
reconstructions.
Our
method
can
be
applied
output
any
estimation
end-to-end
trainable
fashion.
For
evaluation,
manually
annotated
all
test
split
popular
NYUv2-Depth
dataset.
show
that
our
approach
improves
localization
state-of-the-art
could
evaluate,
without
degrading
accuracy
rest
images.
2021 IEEE/CVF International Conference on Computer Vision (ICCV),
Год журнала:
2021,
Номер
unknown
Опубликована: Окт. 1, 2021
As
a
fundamental
building
block
in
computer
vision,
edges
can
be
categorised
into
four
types
according
to
the
discontinuity
surface-Reflectance,
Illumination,
surface-Normal
or
Depth.
While
great
progress
has
been
made
detecting
generic
individual
of
edges,
it
remains
under-explored
comprehensively
study
all
edge
together.
In
this
paper,
we
propose
novel
neural
network
solution,
RINDNet,
jointly
detect
edges.
Taking
consideration
distinct
attributes
each
type
and
relationship
between
them,
RINDNet
learns
effective
representations
for
them
works
three
stages.
stage
I,
uses
common
backbone
extract
features
shared
by
Then
II
branches
prepare
discriminative
corresponding
decoder.
III,
an
independent
decision
head
aggregates
from
previous
stages
predict
initial
results.
Additionally,
attention
module
maps
capture
underlying
relations
these
are
combined
with
results
generate
final
detection
For
training
evaluation,
construct
first
public
benchmark,
BSDS-RIND,
carefully
annotated.
our
experiments,
yields
promising
comparison
state-of-the-art
methods.
Additional
analysis
is
presented
supplementary
material.
2021 IEEE/CVF International Conference on Computer Vision (ICCV),
Год журнала:
2019,
Номер
unknown, С. 10342 - 10351
Опубликована: Окт. 1, 2019
Occlusion
relationship
reasoning
demands
closed
contour
to
express
the
object,
and
orientation
of
each
pixel
describe
order
between
objects.
Current
CNN-based
methods
neglect
two
critical
issues
task:
(1)
simultaneous
existence
relevance
distinction
for
elements,
i.e,
occlusion
edge
orientation;
(2)
inadequate
exploration
features.
For
reasons
above,
we
propose
Occlusion-shared
Feature-separated
Network
(OFNet).
On
one
hand,
considering
orientation,
sub-networks
are
designed
share
cue.
other
whole
network
is
split
into
paths
learn
high
semantic
features
separately.
Moreover,
a
contextual
feature
prediction
extracted,
which
represents
bilateral
cue
foreground
background
areas.
The
then
fused
with
precisely
locate
object
regions.
Finally,
stripe
convolution
further
aggregate
from
surrounding
scenes
edge.
proposed
OFNet
remarkably
advances
state-of-the-art
approaches
on
PIOD
BSDS
ownership
dataset.
arXiv (Cornell University),
Год журнала:
2018,
Номер
unknown
Опубликована: Янв. 1, 2018
This
is
an
opinion
paper
about
the
strengths
and
weaknesses
of
Deep
Nets
for
vision.
They
are
at
heart
enormous
recent
progress
in
artificial
intelligence
growing
importance
cognitive
science
neuroscience.
have
had
many
successes
but
also
several
limitations
there
limited
understanding
their
inner
workings.
At
present
perform
very
well
on
specific
visual
tasks
with
benchmark
datasets
they
much
less
general
purpose,
flexible,
adaptive
than
human
system.
We
argue
that
current
form
unlikely
to
be
able
overcome
fundamental
problem
computer
vision,
namely
how
deal
combinatorial
explosion,
caused
by
complexity
natural
images,
obtain
rich
scenes
achieves.
this
explosion
takes
us
into
a
regime
where
"big
data
not
enough"
we
need
rethink
our
methods
benchmarking
performance
evaluating
vision
algorithms.
stress
that,
as
algorithms
increasingly
used
real
world
applications,
evaluation
merely
academic
exercise
has
important
consequences
world.
It
impractical
review
entire
Net
literature
so
restrict
ourselves
range
topics
references
which
intended
entry
points
literature.
The
views
expressed
own
do
necessarily
represent
those
anybody
else
community.
eNeuro,
Год журнала:
2019,
Номер
6(3), С. ENEURO.0479 - 18.2019
Опубликована: Май 1, 2019
A
crucial
step
in
understanding
visual
input
is
its
organization
into
meaningful
components,
particular
object
contours
and
partially
occluded
background
structures.
This
requires
that
all
are
assigned
to
either
the
foreground
or
(border
ownership
assignment).
While
earlier
studies
showed
neurons
primate
extrastriate
cortex
signal
border
for
simple
geometric
shapes,
recent
show
consistent
coding
also
complex
natural
scenes.
In
order
understand
how
brain
performs
this
task,
we
developed
a
biologically
plausible
recurrent
neural
network
fully
image
computable.
Our
model
uses
local
edge
detector
(
B
)
cells
grouping
G
whose
activity
represents
proto-objects
based
on
integration
of
feature
information.
send
modulatory
feedback
connections
those
caused
their
activation,
making
selective.
We
found
close
agreement
between
our
neurophysiological
results
terms
timing
signals
(BOSs)
as
well
consistency
BOSs
across
benchmarked
Berkeley
Segmentation
Dataset
achieved
performance
comparable
state-of-the-art
computer
vision
approaches.
proposed
provides
insight
cortical
mechanisms
figure-ground
organization.