6. Color
transfer
technique:
limita7ons
Cut-‐and-‐paste
Match
Color
Conflates
the
effects
of
reflectance
and
illumina7on
7. Improvements
of
CTT
(ColorComp)
Lalonde
and
Effros
2007
• Analyze
huge
dataset
of
natural
images
:
difference
in
distribu7on
of
realis7c
and
unrealis7c
images
• Recolor
regions
for
realis7c
composi7ng
Limita7ons:
requires
a
large
dataset,
depends
on
the
presence
of
images
that
are
similar
to
the
target
8. #
of
pixels
Professional
compositors
Shadows
Midtones
Highlights
Brightness
• Isolate
highlights
and
match
their
colors
and
brightness
• Balance
mid-‐tones
with
gamma
correc7on
• Match
the
shadow
regions
9. Color
Harmony
Cohen-‐Or
et
al.
2006
Harmonic
colors
are
sets
of
colors
that
are
aesthe7cally
pleasing
in
terms
of
human
visual
percep7on.
10. Color
Harmony
Cohen-‐Or
et
al.
2006
Limita7ons:
obtained
images
are
not
necessary
realis7c,
ignores
luminance
and
contrast,
the
approach
has
not
been
quan7ta7vely
evaluated
11. Alterna7ve
to
alpha
maAe:
seamlessly
blending
• Feathering
• Laplacian
pyramids
[Odgen
et
al.
1985]
• Gradient-‐domain
composi7ng
[Perez
et
al.
2003]
12. Alterna7ve
to
alpha
maAe:
seamlessly
blending
foreground
background
Cut-‐and-‐paste
Gradient-‐domain
Limita7ons:
2
source
images
should
have
similar
colors
and
textures
13. Problem
statement
• Which
sta7s7cs
control
realism?
• How
do
these
sta7s7cs
affect
human
judgment
of
realism?
• Automa7c
algorithm
to
improve
realism?
14. Good
sta7s7cs
• Highly
correlated
between
foreground
and
background
• Easy
to
adjust
• Independent
from
each
other
15. Categories
of
sta7s7cal
measures
• Luminance
• Color
temperature
(CCT)
• Satura7on
• Local
contrast
• Hue
(circular
sta7s7cs)
16. Sta7s7cal
measures
Standard
devia7on
#
of
pixels
mean
Brightness
• Mean
• Standard
devia7on
• High
• Low
• Kurtosis
• Entropy
17. Find
correla7on
• Pearson
correla7on
coefficient
• Standard
devia7on
of
offset
δi
=
Mif
–
Mib
M
–
measure
f
–
foreground
b
–
background
i
–
sta7s7cs
18. Sta7s7cal
experiment
• Use
large
(4126
images)
labeled
dataset
• Select
the
most
correlated
sta7s7cs
23. Results:
sta7s7cal
experiment
•
luminance,
color
temperature,
satura7on,
local
contrast
are
most
correlated
•
mean
of
zones
correlate
more
than
other
sta7s7cal
measures
•
mean
of
high
and
low
zones
correlate
more
than
mean
of
en7re
histogram
24. Human
subjects
experiment
Experiment
with
human
subjects
on
Amazon
Mechanical
Turk
(MTurk)
• 20
natural
images
• 3
key
sta7s7cs
(luminance,
color
temperature,
satura7on)
28. Automa7c
composite
adjustment
• Zone
selec7on
using
machine
learning
(random
forest
classifier)
T
=
s
*
σg,
s
=
0.1
Three
binary
classifiers:
• Pick
smallest
changes
if
mul7ple
zones
• Features
for
foreground
and
background
and
per
sta7s7c:
std,
skew,
kurt,
entropy,
p1,
p2
…
p20.
29. Pipeline
Input:
foreground
and
background
image
1. Match
H-‐zone
of
contrast
using
S-‐shape
2. Select
zone
and
adjust
mean
of
luminance
3. Select
zone
and
adjust
mean
of
CCT
4. Select
zone
and
adjust
satura7on
Adjust
algorithm
is
greedy,
could
iterate
several
7mes
if
needed