Peter ODonovan University of Toronto Aseem Agarwala Adobe Systems Inc Aaron Hertzmann University of Toronto Choosing colors is hard for many people Choosing colors is hard for many people ID: 309492
Download Presentation The PPT/PDF document "Color Compatibility From Large Datasets" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Color Compatibility From Large Datasets
Peter O’DonovanUniversity of Toronto
Aseem AgarwalaAdobe Systems, Inc.
Aaron Hertzmann
University of TorontoSlide2
Choosing colors is hard for many people Slide3
Choosing colors is hard for many people Slide4
?Slide5
How do designers choose colors?Slide6
PicassoHow do designers choose colors?Slide7
You the DesignerHow do designers choose colors?Slide8
Krause [2002]How do designers choose colors?Slide9
Goethe [1810]Complementary Color Theory: colors opposite on the color wheel are compatibleSlide10
Hue Templates: relative orientations producing compatible colors
Complementary
Monochromatic
Analogous
TriadSlide11
Photo and Video Quality Evaluation:Focusing on the SubjectLuo and Tang 2008Aesthetic Visual Quality Assessment of PaintingsLi and Chen 2009Color Harmonization for VideosSawant and Mitra 2008
Color
Harmonization
Cohen-Or
et al. 2006Slide12Slide13
Adobe
Kuler527,935 themesRatings: 1-5 starsSlide14
Adobe
Kuler527,935 themes
Ratings: 1-5 starsSlide15
Adobe
Kuler527,935 themes
Ratings: 1-5 starsSlide16
Adobe Kuler527,935 themesRatings: 1-5 starsSlide17
COLOURLovers1,672,657 themesViews and “Likes”Slide18
COLOURLovers1,672,657 themesViews and “Likes”Slide19
Goals1. Analysis Test hypotheses and compatibility models2. Learn Models Predict mean ratings for themes3. New Applications Develop new tools for choosing colorsSlide20
Goals1. Analysis Test hypotheses and compatibility models2. Learn Models Predict mean ratings for themes3. New Applications Develop new tools for choosing colorsSlide21
104,426 themes Ratings: 1-5 stars 383,938 themes # Views and “Likes”
Kuler
DatasetCOLOURLovers DatasetSlide22
Mechanical Turk dataset 10,743 themes from Kuler40 ratings per theme1,301 total participantsSlide23
Overall preference for warmer hues and cyansHistogram of hue usageHue% of all themes Slide24
Mean rating for themes containing a hue
Overall preference for warmer hues and
cyans
Hue
Mean RatingSlide25
Histogram of
h
ue adjacency (
Kuler
)Slide26
Histogram of
h
ue adjacency (
Kuler
)Slide27
is more likely than
Histogram of
h
ue adjacency (
Kuler
)Slide28
Significant structureHistogram of hue adjacency (Kuler)Slide29
Significant structureWarm hues pair well with each other
Histogram of hue adjacency (Kuler)Slide30
Significant structureWarm hues pair well with each otherGreens and purples more compatible with themselves
Histogram of hue adjacency (Kuler)Slide31
Hue Template AnalysisSlide32
Hue Templates: relative orientations producing compatible colorsSlide33
Templates are rotationally invariant
H
ue Templates:
relative orientations producing compatible colorsSlide34
Different templates equally compatible
Complementary
Monochromatic
Analogous
Triad
H
ue Templates:
relative orientations producing compatible colorsSlide35
Diagonal lines are hue templates (
Kuler interface bias)
Hue adjacency in a theme (
Kuler
)Slide36
Complementary template
Hue adjacency in a theme (Kuler)
Diagonal lines are hue templates (
Kuler
interface bias)Slide37
Hue adjacency in a theme (
Kuler)Complementary:Slide38
Complementary:
Data:
Hue adjacency in a theme (
Kuler
)Slide39
In template theory, diagonals should be uniform
Hue adjacency in a theme (
Kuler
)Slide40
In template theory, diagonals should be uniform
Large dark bands indicates no rotational invariance
Hue adjacency in a theme (
Kuler
)Slide41
Kuler
CL
Hue adjacency in a theme
COLOURLovers
’ has less
interface
bias
Templates are not presentSlide42
Distance to templateRatingSlide43
Distance to templateThemes near a template score worseRatingSlide44
Themes near a template score worse - “Newbie” factor - “Too simple” factorDistance to templateRatingSlide45
MTurk has no interface bias: much flatterDistance to templateRatingSlide46
Template ConclusionsTemplates do not model color preferencesThemes near a template do not score better than those farther awayNot all templates are equally popularSimple templates preferred (see paper)Slide47
Hue
Entropy: entropy
of hues along the hue wheelSlide48
Hue
Entropy: entropy
of hues along the hue wheel
Low Entropy
Few Distinct ColorsSlide49
Hue
Entropy: entropy
of hues along the hue wheel
Low Entropy
High Entropy
Few Distinct Colors
Many Distinct ColorsSlide50
Hue
Entropy: entropy
of hues along the hue wheel
Hue Entropy
RatingSlide51
Hue
Entropy
:
entropy
of hues along the hue wheel
Hue Entropy
RatingSlide52
Hue Entropy Rating
Hue
Entropy
:
entropy
of hues along the hue wheelSlide53
Hue
Entropy
:
entropy
of hues along the hue wheel
Hue Entropy
RatingSlide54
Main Analysis Results1. Overall preference for warmer hues and cyansSlide55
Main Analysis Results1. Overall preference for warmer hues and cyans2. Strong preferences for certain adjacent colorsSlide56
Main Analysis Results1. Overall preference for warmer hues and cyans2. Strong preferences for certain adjacent colors3. Hue templates a poor model for compatibilitySlide57
Main Analysis Results1. Overall preference for warmer hues and cyans2. Strong preferences for certain adjacent colors3. Hue templates a poor model for compatibility4. People prefer simpler themes (but not too simple)Slide58
Main Analysis Results1. Overall preference for warmer hues and cyans2. Strong preferences for certain adjacent colors3. Hue templates a poor model for compatibility4. People prefer simpler themes (but not too simple)See paper for other testsSlide59
Goals1. Analysis Test hypotheses and compatibility models2. Learn Models Predict mean ratings for themes3. New Applications Develop new tools for choosing colorsSlide60
3.63Slide61
3.63
Slide62
Mean rating over all users
3.63
Slide63
Features (326 total)Colors, sorted colors, differences, min/max, max-in, mean/std dev, PCA features, hue probability, hue
entropyRGB, HSV, CIELab, Kuler color wheel“Kitchen Sink”
3.63Slide64
ModelsConstant baseline: mean of training targets SVM-R, KNN
LassoLinear regression model with L1 norm on weightsSolutions have many zero weights: feature selection
3.63Slide65
Dataset MAEConstant BaselineKNNSVM-RLassoLasso over BaselineKuler0.5720.533
0.5310.5219%
COLORLovers0.7030.674
0.650
0.644
8%
MTurk
0.267
0.205
0.182
0.179
33%Slide66
Dataset MAEConstant BaselineKNNSVM-RLassoLasso over BaselineKuler0.5720.533
0.5310.5219%
COLORLovers0.7030.674
0.650
0.644
8%
MTurk
0.267
0.205
0.182
0.179
33%Slide67
Dataset MAEConstant BaselineKNNSVM-RLassoLasso over BaselineKuler0.5720.533
0.5310.5219%
COLORLovers0.7030.674
0.650
0.644
8%
MTurk
0.267
0.205
0.182
0.179
33%
Many more ratings per theme in MTurkSlide68
Dataset MAEConstant BaselineKNNSVM-RLassoLasso over BaselineKuler0.5720.533
0.5310.5219%
COLORLovers0.7030.674
0.650
0.644
8%
MTurk
0.267
0.205
0.182
0.179
33%
MTurk has an average
std
dev
of
0.33
Kuler
has an average
std
dev
of
0.72Slide69
MTurk Test SetHuman RatingLasso RatingSlide70
High-rated
Slide71
High-rated
Low-rated
Slide72
High-rated
Low-rated
High prediction error
Slide73
Model AnalysisSlide74
Important Lasso FeaturesPositive: high lightness mean & max, mean hue probabilitySlide75
Important Lasso FeaturesPositive: high lightness mean & max, mean hue probabilityNegative: high lightness variance, min hue probabilitySlide76
Goals1. Analysis Test hypotheses and compatibility models2. Learn Models Predict mean ratings for theme3. New Applications Develop new tools for choosing colorsSlide77
1. Improve a ThemeSlide78
Maximize regression
score Stay within a distance
of original (L2 in CIELab)
Slide79
Select order which maximizes scoreSlide80
Optimize colors with CMA [Hansen 1995]Slide81
Original
Best Order
Color and Order
Slide82
Original
Best Order
Color and Order
Slide83
MTurk A/B test
with original and optimized themes
Order and ColorSlide84
2. Choose 5 colors that best ‘represent’ an image Slide85
One approach: k-means clusteringSlide86
One approach: k-means clusteringThis ignores color compatibilitySlide87
Optimize 5 colors thatMatch the image wellMaximize regression scoreSlide88
Optimize 5 colors thatMatch the image wellMaximize regression scoreSee paper for detailsSlide89
With
Compatibility Model
W/O
Compatibility
ModelSlide90
MTurk
A/B
test
with and without
compatibility modelSlide91
3. Given 4 colors for foreground, suggest backgroundSlide92
Given 4 colors, choose 5th color to maximize scoreWant contrast with existing colorsSlide93
Find next best color, away from previous choices
,
, …Slide94
Model Suggestions
Random SuggestionsSlide95
MTurk tests selecting
‘Worst’ and ‘Best’
4 model & 4 randomSlide96
Model Limitations & Future WorkSlide97
Hard to interpretFeaturesWeightsSlide98
Model has very few abstract colors, only 1-D spatial layoutSlide99
VSModel does not understand how colors are usedSlide100
VSSlide101
ConclusionsColor preferences are subjective, but analysis reveals many overall trendsSimple linear models can represent compatibility fairly wellModels can be useful for color selection tasksOur datasets and learned models are available onlineSlide102