Color

From CS294-10 Visualization Sp11

Revision as of 20:00, 28 February 2011 by Manas Mittal (Talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Lecture on Feb 23, 2011

Slides

Contents

Readings

  • Color and information, In Envisioning Information, E. Tufte
  • Charting color from the eye of the beholder. Landa, Fairchild. (html)
  • A rule-based system for assisting color map selection, Bergman, Treinish, Rogowitz, (html)

Optional Readings

  • Color2Gray: Salience-preserving color removal. Gooch, Olsen, Tumblin, Gooch. AMC Transactions on Graphics. (pdf)
  • A framework for transfer color based on the basic color categories. Chang, Saito, Naakajima. (pdf)
  • Color guidelines, Brewer, (html)

Demonstrations

Thomas Schluchter - Feb 22, 2011 09:02:16 pm

Overall, I found the readings for this lecture fascinating. In particular, it hadn't occurred to me that color is such an unstable visual variable the effect of which varies according to both environmental influences (e.g. lighting) and the perceptual abilities of individuals (e.g. aged lenses).

One the one hand, PRAVDAColor is similar to Mackinlay's APT in that it tries to marry an understanding of the data with an understanding of perceptual hierarchies. As PRAVDAColor goes into greater detail with regard to the appropriateness of color choices than Mackinlay's conjectured effectiveness ranking of visual variables, its approach is much more differentiated. (Notice, for example, that in Mackinlay's ranking, the luminance/value criterion is absent.) This is why, on the other hand, PRAVDAColor and APT are very different. The APT treats hue and saturation as distinct visual variables that can take on different dimensions of meaning, whereas PRAVDAColor seems to treat color as a holistic concept with special attention to its various components.

Interesting overall how Tufte's advice is mainly focusing on aesthetics, not so much on perceptual differences. Probably because the examples in his chapter aren't as information-dense as the examples in the other articles.

Julian Limon - Feb 23, 2011 05:43:31 pm

I agree with the previous poster: I also loved this topic. I had had little exposure to color theory before--I had always treated color as a combination of three variables. I also liked the mix of different kinds of readings. While the Tufte reading was about aesthetics and using color with a purpose, the Landa reading was more of the history of the Mussel system, and the Bergman et al. reading was a much more dense and technical one. I'm not sure I was able to grasp everything out of this last reading, but after today's lecture I feel I have a much better sense of how these systems work. It was fascinating to learn the difference between perceptually-motivated and metric-based systems. Thus, the Colorbrewer will definitely be one of my most used resources when trying to come up with a good color palette for a visualization.

It was also really interesting to see the concepts and constraints that were evaluated for Tableau's default palette. The concept of being able to name the color was the most revealing to me. It actually makes a lot of sense once you think of it, but I probably had never though of the ability to name the colors in a visualization. From now on, I'll definitely keep that in mind.

David Wong - Feb 23, 2011 05:54:33 pm

I enjoyed both the reading and lecture for color as I've never had a full walkthrough on how color is defined--from it's physical properties to its perception--and did not know much about the history of the classification of color (e.g. Munsell's system and CIE). Given color's complicated interpretation, I found it interesting that children in Sweden are taught the NCS and even classify their crayon colors using that system. Training children early on in this manner seems like a great idea as it would train them to identify colors in this way, allowing them to have less confusion in describing specific colors.

The lecture described the CIE (x,y,z) categorization of colors, the projection of that color-space onto 2 dimensions, and the available range of colors given certain displays: LED, white-LED, projector, etc. I was wondering how a specific color range mapped light wavelength (from the color space) to specific RGB values (0-255). As each display has different ranges of colors available, I was wondering if a certain RGB combination would appear different given the display, assuming the colors were mapped evenly across the range of available colors.

Finally, the lecture/reading left us with some practical tools. The IBM rule-based approach towards assigning colors to data and Cynthia Brewer's Color Brewer definitely offer good guidelines to follow and pallets to choose from, respectively.

Natalie Jones - Feb 23, 2011 06:33:04 pm

In a design context, I find it interesting how significant our reactions to and perceptions of color can be, yet many people are still pretty ignorant about how to use it. Some of the examples in Tufte's chapter illustrate this well, and in particular his example of old computer programs that used overwhelmingly bright color combinations. They make me think of all the terribly designed websites out there where people seem to feel a need to throw color up on their site in some way, but have no idea how to put it together in a way that facilitates absorption of information (or at the very least, a pleasurable user experience). If we all generally respond to color in the same way, why don't we understand what doesn't work when we see it? It seems like there is some deception at work, in that color in itself can seem like a pleasing thing to look at, and we don't necessarily consider right away whether we're getting any information from it. This could really be a much bigger question that relates to a gap between perception and creation that people in various disciplines are probably studying. For our purposes, though, I'm glad to be made aware of the issues surrounding color and ways to use it effectively in visualizations, because it does seem like something that needs to be pointed out for most people, and will be useful background to have.

Brandon Liu - Feb 23, 2011 07:12:51 pm

I'm interested in the automated color choices made in Tableau in relation to the shapes of marks - most of the ones we've seen have been shapes like circles or boxes. My question is, how does the hue or saturation of marks in a graph affect perceived density? This could have implications for color choices in scatter plots - for example, if darker colors are perceived as denser against a light background, a user may draw different conclusions. I was really interested in some of the slides in lecture on perceptual bias, like the two yellow crosses, but didn't see how they tied in directly to color choices for visualization. One experience I did have was in Assignment 2, when I attempted to draw conclusions about trends on 2D scatterplots based on how dense the colors appeared - I ended up having to fiddle with the size of marks to see meaningful pattens.

Siamak Faridani 21:29, 23 February 2011 (CST)

One thing that is still not clear to me is how to algorithmically generate distinctive colors. In class it was mentioned that colors should be easily named. I am wondering how we can include that in the algorithm. I know that R has this package brewer that performs distinct color selection but I am really puzzled on what is the proper way of doing it.

Michael Hsueh - Feb 24, 2011 01:13:02 am

With regards to David's comment amount mapping RGB values to wavelengths/voltages in a display device and its inevitable variability. This is probably a motivating factor for the existence of all the expensive display calibration software/equipment products. Anyway, something that always strikes me about viewing the CIE diagram with the superimposed device ranges is how small a portion of the entire perceivable color space is actually reproducible in color displays, let alone able to be captured by imaging sensors. I had also wondered about the point Brandon brought up regarding the impact of color on the perception of other visual variables -- provoked partially by the idea of a 'hierarchy' of colors.

PRAVDAColor uses a systematic/algorithmic approach to the important issue of color map selection. It is impressively thorough and nuanced in its treatment of the human visual system. I wonder to what extent such systematic methods have made their way to applications beyond Data Explorer. Then there is the huge realm of 3D visualizations, where perceptual color theories are still more critical as hue, brightness, and chroma themselves additionally encode and interact with spatial elements of the data. Obviously there is much more to color in the context of visualization than most would believe at a glance.

Dan - Feb 24, 2011 10:55:21 am

Edward Tufte provides a great overview of color and its use in visualization. The foundations of its use, whether to categorize information by using color as nouns, show quantities and measures, representations, or simply, beauty, Tufte covers when to use color for each of these cases. I definitely agree with Eduard Imhof's rule about pure bright colors being loud and having "unbearable" effects, however, sometimes when using colors as nouns, having distinct colors can be visually discriminating, whether or not its pleasing. However, I think the point is that you can probably to both. Burham's architectural drawing was nice because it didn't over use color, there was a lot of white space. Tschichold's rejection of centralized layouts is also important. Look at the web, most websites are left justified with pictures in float layouts that span across to one side, instead of being centered. I imagine that there is psychological effects that can confuse a reader (due to the human visual system) when scanning a picture that is centered while reading one of two columns. The reader may not readily know which column to go to after looking at the image.

The Euclidean proofs were great. I've never seen a language where they typeset images with text like that. It was hard to parse, as the graphic is like the legend that you keep having to refer to, however, the over message is very clear and you can follow it algorithmically and the information is received as if a math tutor was proving the euclidean proof right in front of you. I suppose the difference is between having a series of drawings that lead up to the final drawing vs. only the final drawing. I think I would prefer the sequential pictures that build up to the final proof, however.

I also like the ideas applied to user interfaces in the context of information resolution on computer screens. It reminds me of the early days of macintosh when they had the OS 9 black and white windows, they were a bit archaic compared to the beautiful interfaces we see today, which seem much less cluttered, simply by using color instead of lines and patterns. Also, as we have mentioned in the class, color can change to the human visual system when on another color background, hence color isn't reliable in many cases. That said, it's important to pick and choose when to use color and which methods will convey information in a way that is unbiased and provides value to a visualization.

PRAVDAColor was a really robust looking software. Cartographic information, ocean conditions, and many various applications used this software which helped users define parameters to create useful visualizations. Defining a rule-based system for choosing colors and such for visualizations is important, as many people may not have fine arts backgrounds or know how colors interact with the human visual system. Providing a system to assist the user in this process in invaluable. I would like to see this software in action.

Krishna - Feb 24, 2011 02:15:58 pm

Although I liked the PRAVDAColor paper, I still havent got a complete grasp of the rules that are incorporated into the system beyond that of using spatial frequency for selecting values of hue and luminance. One nice feature about the system is that it allows users to interact and fine tune the suggestions made by the rule based engine. An interesting extension of the system would be to consider the changes made by the user and incorporate them as additional rules, these rules can then be used to personalize the colormap associations. Given that color perception is a subjective phenomena, building visualizations that can adapt to user color preferences based on their interaction(feedback) sounds like an interesting idea.

Michael Cohen - Feb 24, 2011 02:43:29 pm

As an aside, it's interesting to note that different species -- and even different humans -- have different numbers of distinct color receptors and therefore see the world differently. It's been a while since I looked into this, but I believe that while many lower lifeforms have three visual axes like we do, birds have four, and most mammals besides primates have only two. They theory on mammals is that the early ones were nocturnal/burrowing and therefore didn't need good color vision. This is why dogs, etc., are color-blind (although they do still have two axes, like most color-blind humans). Primates re-evolved three-axis vision as they began operating more in the daylight, and I believe actually evolved different receptors for the third axis than lower life forms have, leading to somewhat different spectral sensitivities.

Furthermore, there are a couple of variants of one of the color receptors floating around in the pool of human X chromosomes, and some women have one of each variant. The variants respond to slightly different wavelengths, so at least some of these people seem to perceive color on four axes, at least in certain ranges, allowing them to distinguish some spectra that appear as metamers to most of us. So they can actually see more colors!

Sally Ahn - Feb 24, 2011 03:44:47 pm

The PRAVDAColor tool aims to automate some choices for colormaps by taking into account the types of data, representation task, and spatial frequency, which affects how humans perceive luminance and hue. I think their taxonomy of colormaps is a great way to automate color selection. However, I can't help wondering how Tufte would react to some of their examples. For example, the Yearly Average of Global Weather Station Data for 1960 uses bright, saturated colors for large areas on the map and the few boundaries are encoded with duller colors, which violates some of the rules Tufte postulates. It would interesting to see if the system can incorporate some of Tufte's rules and yield better results.

As an interesting side note, here's a rather cool demonstration of simultaneous contrast we saw in lecture: http://www.youtube.com/watch?v=MMJ6thfG2Vg

Matthew Can - Feb 24, 2011 05:58:27 pm

I found the Landa and Fairchild reading on the history of color science to be very enjoyable and informative, especially the part about the Munsell color system. What I like the most about Munsell's system is that the distances along each dimension correspond to perceptual color distances. This is particularly important for visualization, in cases where we encode quantitative variables using color. One of the biggest problems with the RGB system is that distances have no perceptual meaning. One reason for the poor default color schemes for heat and choropleth maps could be that they're chosen based on the RGB system.

In lecture we talked about how our perception of color is heavily influenced by the the context (the viewing environment, surrounding colors, background). This got me thinking about how we can exploit this in computer science. It would be interesting if our interfaces could dynamically adjust their color schemes based on changes to the environment. For example, suppose my desktop background is a light color and my icons are all darkly colored. Now, if I change my background to a dark color, it would be nice if my icons were aware of this and could recolor themselves to provide more contrast against the new background.

Karl He - Feb 24, 2011 11:40:32 pm

The Brewer color schemes seem to implement Tufte's description of the use of color pretty well. It's easy to use color to represent qualitative data, and I feel that's the case that happens the most often. Color representing sequential data is also straightforward, but it's what Brewer labels as "diverging" data that I find it is the most compelling to use color for.

There are many ways to represent qualitative data, and sequential data can be rendered in grayscale fairly easily (although color allows doing qualitative and sequential simultaneously). Using color for diverging data allows two opposite extremes to be emphasized equally, without the bias that would undoubtedly be perceived by someone reading a visualization between black and white. I don't see any other visualization technique being able to duplicate this property nearly as well.

Saung Li - Feb 25, 2011 05:43:48 am

PRAVDAColor looks like a great tool for assisting users in selecting colormaps based on the given data and what the users want from it. A lot of research on various topics, such as perception, cognition, and color theory, has been put into this project to create such a tool, and it seems like it's had success within IBM. I'm curious on whether this tool has been more widely used by other companies, as this tool can help lots of people in creating visualizations.

I liked the discussions from lecture overall, as they gave me a better sense of how to use color in visualization. The simultaneous contrast examples show how foreground colors can be influenced by background colors, and the examples on when to use saturation, hue, lightness is helpful in letting viewers understand images better. The ninja turtle example is most fascinating, as it shows how our eyes can adjust to particular colors.

Manas Mittal - Feb 28, 2011 01:43:13 pm

I thought the American Scientist work as among the most interesting pieces I've read this month. I liked the historical context provided in the article, and the description of the Munsell's work. In reading the Bergman paper, I was recently reminded of the default color scheme used by the Google Charting API, which uses chroma variation rather than hue variation.

I was wondering if we can combine the MacKinlay's work and distill this to include mappings for particular visualization types. For example, pie charts work well with representing the division of a larger 'pie', and for those, it would make sense to use hue rather than chroma.



[add comment]
Personal tools