1. TVCG 2013 Sungkil Lee, Mike Sips, and Hans-Peter Seidel

2.  Introduction  Class Visibility  Optimization  Example  Conclusion

3.  Principles of effective color palettes (Trumbo, 1981) ◦ Order: colors chosen to present an ordered statistical variables should be perceived as preserving that order. ◦ Separation: colors chosen to present the differences of a variable should be clearly perceived as different.

4.  ColorBrewer.org: An Online Tool for Selecting Colour Schemes for Maps (The Cartographic Journal 2003)  http://colorbrewer2.org/

5.  In this work, it focus on separation, which is still challenging. ◦ Wrong use of categorical colors may lead misleading interpretation of data. ◦ Many visualization packages provide little guidance on color selection.

6.  Dominant structures locally suppress small inhomogeneous groups. ◦ Even the same colors are not perceived as equal and distinct.

7.  Dominant structures locally suppress small inhomogeneous groups. ◦ Global contrast enhancement alleviates the symptom, but not that much.

8.  A measure of perceptual intensity for categorical differences.  An optimization metric to make categorical differences visible to users.

9.  Perceptually-driven: ◦ class visibility as a metric to measure the perceptual intensity of a categorical groups  Visibility equalization by optimization: ◦ to optimize the class visibility function of input data ◦ The optimization aims to balance the local contrast of the image between small and large groups.

11.  Hence, we instead introduce point saliency, which exactly weight the contribution of categorical color at a single data point. ◦ The point saliency is measured as the color difference from the average color of neighboring data points.  Color difference, Δ, is measured by the CIE76 metric, which is the Euclidean distance in CIE Lab.

14.  The class visibility, defined for each data point, is estimated by: ◦ weighting its point saliency with its structural weight the number of data points, in a neighboring surround, whose labels are the same as the center point

16.  The mean class visibility of each class is computed by the average of the data points that belong to its category.

17. Note that hue values (θ ) will stay constant during the optimization. Target Visibility T: the mean over all V m L k and r k are the lightness and chroma of C k in LCH space.

18. barrier function:

19. i and j indicate the color pair selected from all the classes J : just noticeable difference (JND) of Lab color (3-7 JND in this paper)

20. The final resulting cost functional with the constraints to optimize is: In practice, they found that a fixed value of 10 produces satisfactory results anyway. nonlinear conjugate gradient-based minimization

24.  Our optimization attempts to equalize all the mean class visibilities to a single target constant T.  T mean : average of mean class visibilities ◦ Strong classes are weakened, while weak classes are improved.  T max : maximum among all the mean class visibilities ◦ All the classes are strengthened.

27.  class visibility as a measure of perceptual intensity in categorical data visualization  an automated color optimization algorithm to make categorical distinction clearly visible  No simple rule to choose target visibility: T mean, T max, or others.  During the optimization, hue value was fixed. ◦ the change of hue will be useful for large-scale discrimination