Validating large-scale lexical color resources

Validating large-scale lexical colorresourcesGiordano Beretta & Nathan Moroney, Hewlett-PackardDate: 8/June/2011

Color naming #EF4123 #007CB0•  Colorimetric values are best for #848688 communicating color via machine #BF1E74 #007CB0•  Color terms are best for communicating color among humans #008F4C•  Problem: how can we ﬁnd the most effective color terms? #F89F6D #F499B82 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

World Color Survey, Berlin & Kay, 1969•  Munsell Sheets of Color are shown to respondents to elicit color terms•  A snapshot in time•  Experiment in the wild•  Several similar experiments, e.g., ISCC–NBS3 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Light + objectThe spectral power distribution of the light reﬂected to the eye by an object is theproduct, at each wavelength, of the objects spectral reﬂectance value by the spectralpower distribution of the light source CWF Complexion 400 500 600 700 400 500 600 700 400 500 600 700 Incident SPD x Reflectance curve = Reflected SPD Deluxe Complexion CWF4 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without 400 notice. 500 600 700 400 500 600 700 400 500 600 700 Created 26/4/2011

Is it robust? g blue green• Experiment by Boynton & Olson proves robustness w.r.t. light source Robert M. Boynton, Insights gained from naming grey white the OSA colors, Color categories in thought and yellow language (Clyde L. Hardin and Luisa Mafﬁ, eds.), purple brown Cambridge University Press, 1997, pp. 135–150. pink red orange j5 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Emissive color?• Excellent correlation between controlled reﬂection experiment and uncontrolled crowd-sourced experiment on the Web Giordano B. Beretta and Nathan M. Moroney, Is it turquoise + fuchsia = purple or is it turquoise + fuchsia = blue? , vol. 7866, SPIE, January 2011, p. 78660H.6 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Advantage of Web experiments• Crowd-sourcing uses the World Wide Web to recruit thousands of respondents• Persistence in time can account for ephemerality of color terms• Respondents recruited mainly from the color community• Nathan Moroney, Dimitris Mylonas7 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

We do not have so many friendsCan we leverage a larger class of respondents?•  Munroe and Ellis of xkcd fame have performed a color naming experiment among their readers•  Is there any scientiﬁc value in such totally uncontrolled data?•  Frequency sorted color term data points contributed (TDP):8 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Validation experimentCan we leverage a larger class of respondents?•  Multi-stimulus categorization task•  Strict sRGB conditions•  16 color normal observers•  Instructions: “select the color patches you might use with the color term”•  Contributed color stimuli per term (TDP without frequency), CCS:9 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Complete Munroe & Ellis versus the Berlin & Kay data Graphs show the Berlin and Kay averaged centroids on the abscissa and the Munroe and Ellis averages on the ordinate. h*ab L* C* y = 0.9371x + 8.6746 y = 0.9122x + 6.5418Munroe & Ellis: All Data Munroe & Ellis- All Data Munroe & Ellis- All Data 360 R2 = 0.97 90 80 R2 = 0.67 270 70 60 180 50 40 90 y = 0.5291x + 41.072 R2 = 0.42 0 30 20 0 90 180 270 360 30 50 70 90 20 40 60 80 Berlin & Kay Berlin & Kay Berlin & Kay 10 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Validated Munroe & Ellis versus the Berlin & KayColor data retained after validation: Brown Purple Pink Orange Blue Yellow Red Green TDP 33% 44% 49% 49% 53% 70% 73% 74% CCS 2% 7% 6% 3% 7% 3% 5% 18% h*ab L* C* 100 y = 0.9286x + 10.427 y = 1.0863x - 0.6702 360 Munroe & Ellis- Validated Data Munroe & Ellis- Validated Data Munroe & Ellis- Validated Data R2 = 0.97 90 R2 = 0.65 80 270 70 60 180 50 40 90 y = 0.3221x + 62.678 R2 = 0.13 0 30 20 0 90 180 270 360 30 50 70 90 20 40 60 80 Berlin & Kay Berlin & Kay Berlin & Kay 11 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Complete versus validated Munroe and Ellis data h*ab L* C* 360 y = 0.9933x + 1.4356 y = 1.2005x - 8.9949 y = 0.9839x + 12.971 R2 = 0.999 90 R2 = 0.99 90 Munroe & Ellis- Validated Data Munore & Ellis- Validated Data Munroe & Ellis- Validated Data R2 = 0.82 270 70 70 180 50 50 90 0 30 30 0 90 180 270 360 30 50 70 90 30 50 70 90 Munroe & Ellis- All Data Munore & Ellis- All Data Munroe & Ellis- All Data 12 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

Conclusions•  Color terms are well suited for human communication•  Color naming experiments are robust w.r.t. •  light sources •  reﬂection vs. emissive patches •  crowd-sourcing •  disruptive users in large experiments•  Color terms are ephemeral and need continuous experiments•  Future work: •  basic terms vs. long tail •  how does color naming scale?13 of 13 © Copyright 2011 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Created 26/4/2011

http://www.mostlycolor.ch	166
http://feeds.feedburner.com	3
http://www.slideshare.net	1
url_unknown	1
http://translate.googleusercontent.com	1
http://webcache.googleusercontent.com	1

Validating large-scale lexical color resources

by Giordano Beretta on May 20, 2011

Accessibility

Categories

Upload Details

Usage Rights

6 Embeds 173

Statistics

Validating large-scale lexical color resources Presentation Transcript