Presented at Color Imaging XVIII: Displaying, Processing, Hardcopy, and Applications in 2013. Application of machine color naming to 200,000+ wikipedia images.
1. Color naming 65,274,705,768 pixels
Nathan Moroney and Giordano Beretta
HP Labs
Electronic Imaging 2013: Color Imaging XVIII
2. Outline
Motivation
More (pixel) data
Finding and processing 65 billion pixels
Hint: Wikipedia & a dual core Open MP color namer
What did you learn?
The most frequent non-achromatic color term is…
What’s next?
Other than a trillion pixels
Electronic Imaging 2013: Color Imaging XVIII
3. Motivation
Previous work in crowd-sourcing color training data
and experimental efforts
Related work in the area of big (image) data
A. Torralba, R. Fergus, W. T. Freeman, "80 million tiny images: a
large dataset for non-parametric object and scene recognition",
IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol.30(11), pp. 1958-1970, 2008.
Ben Shneiderman, "Extreme Visualization: Squeezing a Billion
Records into a Million Pixels", SIGMOD Conference, pp. 3-12,
(2008).
Steven Seitz, “A Trillion Photos”, EI’13 Keynote (2013).
Electronic Imaging 2013: Color Imaging XVIII
4. Motivation
0 1 2 3 4 5 6
Log Number of Images
Electronic Imaging 2013: Color Imaging XVIII
5. Source Data
ImageClef 2010 snapshot
Adrian Popescu, Theodora Tsikrika and Jana Kludas, "Overview
of the wikipedia retrieval task at ImageCLEF 2010", In the
Working Notes for the CLEF 2010 Workshop, 20-23 September,
Padova, Italy, 2010.
250,000 images plus associated wikipedia data
20 gigabytes
65,000,000,000 pixels uncompressed
Electronic Imaging 2013: Color Imaging XVIII
6. Source Data: At 200 PPI
Electronic Imaging 2013: Color Imaging XVIII
7. Processing
Basic single dual-core (but Open MP threaded) script
to process over all image files
Simple stuff like getting image dimensions can be
done over lunch
Uncompressing all the JPEG files to memory can
take hours
Goal was a color naming algorithm that could be run
in less than a day
Electronic Imaging 2013: Color Imaging XVIII
8. Processing
Some testing done using HP Cloud Services and
compute clusters
But majority of focus on single computing device
Antony Rowstron, Dushyanth Narayanan, Austin Donnelly, Greg
O'Shea, and Andrew Douglas. "Nobody ever got fired for using
hadoop on a cluster", In HotCDP 2012 - 1st International
Workshop on Hot Topics in Cloud Data Processing, (2012).
Electronic Imaging 2013: Color Imaging XVIII
9. Processing
Won’t describe the specifics of the color naming
algorithm (throw produce if you have it) but generally
Input single RGB pixel and output is a single color term
Size of vocabulary or number of color terms is a parameter
Relative range of chroma values corresponding to an achromatic
values is also a parameter
Also currently testing a completely revised model
Finally, in the Future directions section note that the
best option for formal publication is to make use of
currently available open source machine learning
toolboxes.
Electronic Imaging 2013: Color Imaging XVIII
10. Results: Aspect Ratios
Wide range of
image types
Most basic test
of processing
scripts
Electronic Imaging 2013: Color Imaging XVIII
11. Results: Median
Additional test and
visualization of
basic color
properties of images
Large enough data
set was worthwhile
to write custom
HTML5 2d canvas
renderer
Electronic Imaging 2013: Color Imaging XVIII
12. Results: Median
So much data, that
as noted by
Shneiderman the
density plot "uses a
spatial substrate
organizing
principle, but shows
concentrations of
markers” is maybe a
better idea
Data, alpha=0.05
Electronic Imaging 2013: Color Imaging XVIII
13. Results: Max
Max of R+G+B for
the images
Final test of basic
scripting code
Electronic Imaging 2013: Color Imaging XVIII
14. Results
Color terms
across all images
Majority pixels
achromatic
Top chromatic
colors are
arguably natural
tones
Higher chroma
terms relatively
infrequent
Electronic Imaging 2013: Color Imaging XVIII
15. Results
Color Terms for 200,000+ images
60000
Color terms per
image 50000
Peak at 5 are all 40000
achromatic terms
Number of Images
30000
or images
Gradual then 20000
rapid usage of 10000
chromatic terms
0
0 5 10 15 20 25 30 35
Number of Color Terms. Maximum Vocabulary of 30
Electronic Imaging 2013: Color Imaging XVIII
16. Results
Color Terms for 200,000+ images
60000
Sudden drop off
at 30 is a model 50000
failure 40000
Term added to
Number of Images
30000
vocabulary based
on previous 20000
limited
10000
optimization
0
0 5 10 15 20 25 30 35
Number of Color Terms. Maximum Vocabulary of 30
Electronic Imaging 2013: Color Imaging XVIII
17. Current Work
Repeated entire process adjusting the model
parameters
Processing to fill SQL databases
Query the database to validate all of the steps and
explore specific
Electronic Imaging 2013: Color Imaging XVIII
18. Current Work
SELECT * from
cntable order by
skyblue desc limit 40
Electronic Imaging 2013: Color Imaging XVIII
19. Future Directions
Image collections as “pixel
corpora” for algorithm
design, testing and optimization.
Similar to the role that written and spoken
corpora fill for NLP and corpus linguistics
Useful to formalize for citation and
repeatability
Additional analysis features
Testing with more public domain
machine learning algorithms for
repeatability
Electronic Imaging 2013: Color Imaging XVIII
20. Summary
Algorithm optimization, like machine color
naming, with 200,000 images is different than with
200.
Based on Wikipedia, majority of visual content or
pixels are achromatic
Based on Wikipedia, higher chroma named pixels are
less frequent
Based on Wikipedia, there is a gradual then sudden
transition in color term usage
Electronic Imaging 2013: Color Imaging XVIII