Wednesday, July 21, 2010

The Calibrated Lunch

The calibrated lunch is one part ice-breaker, one part quasi-experiment and one part an exploration of the self-selection of non-basic color terms in the context of a corporate social dining event.

To start you'll need individually calibrated and uniquely numbered printed color patches. Then you'll need a couple hundred hungry co-workers. Beer is optional.

But let's start the calibrated lunch from the beginning. We have periodic lunch socials at work and each social is hosted by several of the labs. Our lab was to be one of the host labs for an upcoming lunch and after a lunch bike ride John asked if we might do some sort experiment as part of our lunch. Skip ahead a few weeks to a passionate discussion at Antonio's with John and Giordano and the basic idea was formulated:

What if we gave our co-workers randomly colored patches and had the tables with color terms on them - how would people select their tables?

With Norma and Nina's encouragement the idea was refined, tested out on a smaller group of volunteers and orchestrated. This would probably be a good time to thank John, Giordano, Norma, Nina, Gary, Eric, Prith, Paul, Wei, Doris, Wenjia, Laurie, Bob, Ido, Miheer, Rich, Dick, Bill and others that I may have unintentionally forgotten for their help and support. For the purposes of brevity, introductory tangential digressions on praxis, fluxus and performative science have been condensed into this single sentence.

To start with 25 non-basic color terms were selected and psuedo-random noise was added to the corresponding centroids. This is harder than it sounds. The terms purposely excluded the basic color terms of red, green, blue, yellow, black, white, gray, purple, pink and brown. Otherwise the game would be too easy. The non-basic terms were selected to both cover a wide range in color space and to have some overlap between the tables or color samplings. The centroids were taken from the color thesaurus. A visualization of the original color chips is shown below.

Next the colored patches were assigned a unique psuedo-random identifier and printed on a high gloss media with an ink-jet printer. The spectral reflectance curves were then measured for each of these patches. A preliminary calculation of the mean color difference from the mean in CIELAB was used to determine that the above sampling had roughly equivalent variation per centroid. Caveats apply but the point is that each patch was measured and with a measurement of the illumination conditions of the cafeteria the calibration of the lunch is complete.

After the patches were printed and measured, the physical tables were arranged. We set up 25 tables of 10 chairs each in a radial configuration. Each table had a balloon with a non-basic color term on it.

Printed instructions were placed on each table and the instructions were also given verbally while people waited in line for food. The instructions were as follows:

  • Please take a colored patch from one of the volunteers before getting your food.
  • The game consists of sitting at the table with color name closest to your colored chip.
  • If you have any difficulties with selecting a table please ask for help from anyone.
  • The table with closest agreement or smallest difference between their color patches will win a prize.
  • One member of the winning table will then be eligible to win the grand prize.

    Note that it was in fact a competitive calibrated lunch. At this point we will skip over the observed individual and group dynamics of the table selection and optimization. This is arguably a key part of this quasi-experiment but most likely insights gained here will show up in future experiments, posts and analysis. Suffice to say people seemed to be enjoying themselves. Now on to the results. A computational test was done to remove obvious outliers (such as the two pumpkins that sat at the teal table). Interestingly the number of outliers was quite low, around 4% of the participants. After outlier removal the final color seating arrangement was the following:

    After some weighted and non-weighted calculations of the mean color difference from the mean in CIELAB the lime green and thistle tables were announced as winners. Horay for lime green and thistle! Now more caveats, yes Ramin using CIELAB is debateable and yes Gary V. the use of the mean is questionable. But we can do subsequent analysis to see what difference using modern color difference metrics and non-paremetric statistics would have had on the results. Enough custom code had already been written for the calibrated lunch.

    So how do the results for the calibrated lunch compare to the uncalibrated web?

    The CIECAM02 hue angles for the non-basic color terms were computed for both the final configuration of measured printed color patches for the calibrated lunch and plotted versus the corresponding uncalibrated, but assumed to be sRGB, web color centroids taken from the online color naming experiment. The results are shown below and the r-squared is 0.99.

    After excitedly handing a copy of the above plot to Giordano, he summarized the results as demonstrating that our calibrated lunch was 'crowd-compliant'. Indeed. A stamp with a crowd-compliant logo would be quite nice...