Perception

From CS294-10 Visualization Fa07

Jump to: navigation, search

Lecture on Sep 12, 2007

Slides

Contents

[edit] Readings

  • Perception in visualization. Healey. (html)
  • Graphical perception. Cleveland & McGill. (jstor)
  • Chapter 3: Layering and Separation, In Envisioning Information. Tufte.

Optional Readings

  • Gestalt and composition. In Course #13, SIGGRAPH 2002. Durand. (1-up pdf) (6-up pdf)
  • The psychophysics of sensory function. Stevens. (pdf)

[edit] Hazel Onsrud - Sep 10, 2007 12:13:19 am

I thought Tufte's opening sentence in chpt. 3 of Envisioning Information made a good point. Paraphrased: Data may be confusing and cluttered, but information is not, and consequentially visualizations should not be as they are to convey information. Sometimes I feel people too often blame their results on the data...but the information they are gaining from the data ought to be clear and accurate or not used at all.

Also, I liked the 1+1 equals 3 or more rule and wonder how this rule changes in three dimensions. Extending this into 3D it seems 1+1 automatically equals more as the amount of negative space in 3D would be amplified. This also may imply that clutter seems more likely in 3D.

[edit] Omar - Sep 16, 2007 09:02:50 am

at the end of last class, maneesh asked us to ponder: why pie charts? it'd be interesting to have an experiment to try and tease out why people choose pies. here are some hypotheses:

- tools

first i wondered if excel somehow encouraged pie charts. it doesn't seem to, though when choosing a chart type they all kind of look the same except the pie, and that might encourage the user to choose the pie (visual difference).

- culture

most of us have likely grown up with references to circle metaphors, like when referring to money to split: "give him half the pie." in school, i recall learning fractions by using pie-based reasoning (and even reference to apple pie, pumpkin pie). if one encounters a sliced pizza, we're often doing the same kinds of calculations. what's interesting is that in normal life the pies we encounter usually have uniformly sized slices, which isn't the case for pie charts.

- looks/difficulty

there's something just visually appealing about a circle, as compared to a bar chart. maybe that's cultural. in addition, i feel like i may be conveying that i did something more taxing if i use a pie chart, as opposed to a bar chart (though i'm not sure if this is a valid feeling anymore).

- constant area

if you have a bunch of data to divide up, i feel like the natural tendency with bar charts is to make them bigger (add another column) whereas with pie charts you can still use the same area, and just adjust the slice sizes. of course, the pie chart just becomes more difficult to interpret, but perhaps people think this way when making their selection.

on cleveland and mcgill and patch maps: wow, i shouldn't have done a patch map for my redesign :) one thing i will say is that even though they try to address the clustering issue by progressively changing the rectangle sizes in their murder-rate framed rectangle map, all the white space makes it somewhat more difficult to perform a clustering,visually. now, i expect the authors would counter that we might be doing an inexact clustering when using shading, and that may be true, but if the goal is to quickly identify trends and then dig deeper, perhaps the patch map is still useful. i also wonder if there's any good solution (patch or rectangle) that can work with small countries where it's very difficult to fit the rectangle, and the patch largely goes unnoticed (though the patch will work better if there's a jump in that area).

[edit] Athulan - Sep 16, 2007 08:34:39 pm

I'd like to add to Omar's comments. I think that there is a certain elegance with the constant area of the pie-chart. The representation is compact as the size of the pie can be held constant; of course, as Omar pointed out this may make it hard to read with too many slices, but since we already know the total size, the total space we allocate for the pie is set. Also, the relative position of intervals seems to be of less importance in a pie-chart - I have not seen any one convention being enforced in pie-charts for the point (r=R, theta=0). The pie-chart could also be more appealing as it always stays symmetric, and this somehow makes it easier to read.

However, the cultural reasons for liking pie-charts do not apply to me as I grew up in India and I was not much acquainted with pies. I did eat circular-shaped foods, but they were not divided in slices - but I can see why pie-charts would seem natural if you grew up with pies.

On a related note, this website claims that Florence Nightingale invented the first pie-chart

[edit] Ken-ichi - Sep 17, 2007 12:03:35 am

I enjoyed most of Healey's paper. I think the idea of measuring what kind of marks and differences get detected "preattentively" is a cool and useful way to objectively guide how you design info visualizations, and the change blindness demos were super fun. For the first one (plane engine) I actually didn't notice the difference until I not only slowed it down, but skipped the blank frames all together. I showed it to my brother without telling him that, and reported exactly the same thing. Fun.

What I didn't necessarily understand or buy were the models of visual cognition. They seemed to arbitrarily derived from observation, and I was having trouble finding ways to falsify any of them. I was really intrigued by the brief final mention of nonphotorealism. I'm sort of obsessed with field guides (for identifying animals and plants), and there are always people who argue for or against illustrations (flexible, ideal, focused) or photos (accurate, detailed, real) in field guides. I wonder if anyone's tested different techniques for field ID, or if anyone's tested hybrid approaches (like your shape and detail enhancement work, Maneesh).

[edit] Charlotte Wickham - Sep 17, 2007 09:06:55 pm

A thought on the pie: we seem to be very good at recognising standard fractions of a pie like 1/2, 1/3 and a 1/4 (my hypothesis for this is our frequent measurements of this nature on analog clocks). I think data that boils down to close to these fractions (or where segments add to these fractions) can be quite successful as pie charts (although if there are four segments each of 1/4 a graphical representation isn't really necessary).

[edit] Mark Howison - Sep 18, 2007 05:47:16 pm

To add to Charlotte's comment:

I agree that identifying simple fractions of a pie, such as 1/4 and 1/2, is probably an easy and accurate task due to most people's experience with analog clocks. So a pie chart that needs to display a set of data such as {1/2,1/3,1/6} might be fairly successful. Compared to a bar chart of the same data, the pie chart may even better communicate the proportions of the whole. However, now think about a situation where there are lots of pieces of a pie that are varying in size, are not simple fractions, and may have other pieces that are close but not exactly the same in value. For instance, here is a pie chart of the federal budget that I dug up on Google image search [1]:


Image:FederalBudgetPieChart.jpg


What is the percentage going toward job training? It is hard for me to estimate because I have to take a piece of the pie I can recognize, say 1/4 of the pie, and see how many "job training" slices fit inside of it. With a bar chart, I can just read horizontally over to the value on the y-axis.

Also, are the percentages for housing and science/energy the same? With a bar chart, I could visualize a horizontal line across the tops of the bars, and try to judge if one of the bars is shorter or longer.

I think Cleveland and McGill's findings would indicate that a bar chart of this information would allow for more accurate estimation of values and comparisons between values.

[edit] Jimmy - Sep 18, 2007 06:01:43 pm

The preattentive processing in Healey's paper is very interesting. The graphic examples explain how fast our eyes can detect unique visual property in the target, without searching serially each display to confirm the target's presence or absence. Treismen's research is also interesting. She separates the visualization in individual feature maps, so the target detection in each feature can be done in parallel. The problem is the conjunction feature, which we often need to scan serially to detect the combined features. However, she points out that sometimes the conjunction detection can be preattentive, if there's significant difference between target-nontarget features. So we can first filter out the non-target feature, such as the color, and then we can detect the target in the reduced set. But I think in this case it might not be really preattentive, as I it takes a lot more reaction time than individual feature detection. For instance, we tend to search serially in the conjunction search example shown in Fig. 3, which is to detect red circles within red squares and blue circles. According to Treismen's hypothesis, we could simply ignore the blue circles and detect red circles within the red ones. However, Healey said that many studies indicate people can't preattentively process it.

[edit] Robin Held - Sep 18, 2007 08:33:37 pm

I enjoyed the Healey reading the most this week. He wrote succinctly and covered a good deal of ground. In particular, the portion on change blindness was very interesting. I would like to hear Healey and Tufte's opinions on how the concept of change blindness can be applied to the use of small multiples. It seems to underscore the notion that small multiples must be simple and very easy to compare. In particular, the reader's attention must be drawn to the comparisons the author is trying to convey, so any important changes between the multiples aren't lost to change blindness.

[edit] James Andrews - Sep 18, 2007 08:59:10 pm

I'm really not convinced that the pie chart is that bad -- if there are data points which are too close together to be clearly ordered from the pie chart, and ordering them is important, then sure, don't use it. But it seems to convey the context of the data (that it is a set of percentages that add to a whole) much more naturally than a bar chart, and for purposes of data presentation (not exploration) that can be the most important thing.

Also, wikipedia linked to an interesting journal article ([2]) which includes an overview of the studies related to the chart, and it really doesn't seem that the experimental evidence against the chart has ever been very strong ... so even if people are paying attention to the literature it makes sense that they would keep using pie charts. That could answer the 'why do people use pie charts?' question pretty well!

To summarize the history in that paper:

  1. The first widely circulated textbook on statistical graphs, in 1914, was rather down on the pie chart.
  2. The first psychological experiments on graphs, where subjects were required to estimate quantities represented in graphic form, were rather positive -- Eells in 26 showed that subjects could estimate the size of a proportion more quickly and accurately with pie charts than with bar charts.
  3. Studies for many subsequent years were inconclusive. Some studies in 1955 and 59 demonstrated that the pie was not inferior to the divided bar when users had to estimate or compare simple proportions.
  4. In a widely read and influential review, Macdonald-Ross in 1977 concluded that the bar chart was superior to the pie chart based on his reading of the problematic past literature.
  5. Cleveland and McGill in 1984 developed their theory of graphical perception, which implied that, because judgments of angle are generally less accurate than judgments of length, pie charts are worse than bar charts for proportion estimates and comparisons.
  6. Simkin and Hastie in 1987 pointed out that the task used by Cleveland and McGill was not a fair comparison because the key judgment of pie charts is the comparison of a segment angle to the whole 360 degrees, not the difference between two internal angles. They showed by experiment that pie charts and bar charts give equivalent performance when the task is to estimate what proportion of the whole a bar or segment occupies.
  7. Spence showed in 1990 that proportions in pies were judged more accurately than with several other basic forms, and then in 91 with Lewandowsky found that pies and bars performed comparably for simple judgments of comparison but that pies are better when more complicated judgments were required (for example compound comparisons like A+B vs C+D)
  8. Hollands and Spence in 92, 98, and 01 performed more experiments showing that the pie chart was as good as other charts at displaying the relative size of a small number of proportions.

Spence also writes: "In my opinion, much of the adverse criticism of the pie has come from those who have wished it to do more than it could. The pie chart is a simple information graphic whose principal purpose is to show the relationship of a part to the whole. It is, by and large, the wrong choice as an exploratory device, and it is certainly not the correct choice when the graph maker or graph reader has a complicated purpose in mind, such as displaying small changes in proportion over time, a task that would require several pies."

[edit] Omar - Sep 18, 2007 10:40:33 pm

james -- interesting find on the journal article! regarding #6: i think cleveland and mcgill's study wasn't the wrong comparison, as indeed we very often compare sections of a pie chart. think of budget comparisons (did education get more or less than health, how much, etc..). those are not comparisons to the whole. and if they found that performance when comparing to the whole is similar to bar charts, then with c&m's result that just gives more support to 'why not just use a bar chart?'

i think a somewhat convincing point is the 91 result for A+B vs C+D. i can see that perhaps being easier with a pie chart than a bar. but i imagine that's rarely the intent of user of a pie chart? spence's quote is reasonable -- perhaps people do want more when all it can really do is part to whole. but i think another question is 'why not use a bar chart?'

[edit] David Jacobs - Sep 18, 2007 11:19:27 pm

Well, it's kind of hard to defend a pie chart against a simple table of numbers (since any time a pie chart is effective, it has to have few enough slices to keep the angles resolvable), but I think pie charts do have a definite advantage over divided bar charts, or even regular bar charts. This advantage lies in the implicit maximum value. Each slice of the pie represents a percentage inherently, without the need to draw a measurement axis. So technically pie charts have a perfect data ink ratio.

Also, there is something to be said for the design of a pie chart. Pie charts use angle, area, and length redundantly to represent the percentage each slice represents. Each scales proportionally with the value of the slice. This makes it much faster to get the gist of the data, whether or not fine comparisons are possible. We could even add more redundancy by say setting the luminance of each slice according to it's percentage (make smaller slices brighter, to make them more visible against the background large slices). Also, because there is no baseline or scale required for pie charts, it's almost impossible lie with one, which is kind of nice.

[edit] James O'Shea - Sep 19, 2007 07:51:36 am

I was particularly interested in Healy's discussion of 3D perception with respect to texture synthesis and some of the recent non-photorealistic rendering (NPR) approaches. Shading is another cue that is being explored as a way to affect 3D shape perception, and there have been several recent NPR methods developed. One of the main approaches is to emphasize particular surface features by adopting a lighting model that is physically unrealizable. An example of this is described in this paper by Simon Rusinkiewicz.

I also wanted to respond to Ken-ichi's comment above. First of all, I agree that the question of whether to populate a field guide with illustrations or photos is an interesting one. I think informed illustrations can be far more effective in conveying important information than photographs alone. My personal experience is with bird guides, and I think the recently released Sibley guides (and his high quality watercolors) are much better than any of the other guides I have that use photographs. You also asked whether anyone has tried to evaluate some of these NPR methods. I don't think anyone has started to look at Maneesh's multi-light image collections yet, but he and I are working on a project to evaluate the exaggerated shading paper I mention above. Specifically we are conducting psychophysical experiments to study shape perception using various shading conditions. Doug DeCarlo is another researcher studying these types of NPR algorithms.

[edit] Amanda Alvarez - Sep 19, 2007 08:17:55 pm

I'm surprised (although as a vision scientist I shouldn't be) by the disagreement between the various rankings of elementary visual elements. Even the numerous experiments (eg. outlined by Healey) can't create agreement as to which visual things we perceive first and most easily. Mackinlay's rankings are rather long, containing some elements I would think of as second-order (density or connection), and changing from one data type to the next. Cleveland & McGill develop the question a little by asking which variables lead to perceptual accuracy, not just speed, but they don't justify their hierarchy beyond the top three positions. For Healey, the ease or speed of visual variables is summed up in the word 'preattentive', and he provides a whole list of which variables that fall into this group (motion and 3D orientation are very compelling and 'pop out' but I still would classify them as second or third-order because a lot of background calculations go into them; this leads me to believe that perhaps the most effective visual variable is the one that contains the highest or most sophisticated degree of information, yet still retains its 'preattentive' quality). Conversely, we have the sparse classification of Julesz and his textons; incidentally these fit in rather nicely with the elementary operations we believe the brain does (edge detection and so forth). I was going to comment about how Cleveland & McGill did not deal with 'position' and how it interacts with proximity, but finally toward the end of the paper they addressed this: "The position-length experiment suggests, however, that a revision in the theory might be appropriate. Although Judgment Types 1-3 involved judgments of position along a common scale, namely the vertical scale of the bar charts, the horizontal distance between the graphical elements being judged varied from 0 cm for Type 1 to 2.8 cm for Type 2 to 5.6 cm for Type 3; Figures 16 and 17 show that errors increased in going from Type 1 to Type 2 to Type 3. This suggests that the elementary task of judging position be expanded into a continuum of tasks for which accuracy is conjectured to decrease with increasing distance between the graphical elements encoding the data, where distance is measured perpendicular to the axis along which the data are plotted. Not surprisingly, after just two experiments a revision in the theory appears necessary."

In the end I am inclined to throw my hat in with the Gestaltists because they are nicely hand-wavey, and take things like symmetry into account (something which I did not see in any of the other rankings). Also, I don't believe in attention.

[edit] N8agrin - Sep 26, 2007 10:03:44 am

The Healey paper really took me by surprise. All of the examples which are presented, especially the interactive examples in the online version, are a startling insight into how the vision system processes images. This information suggests the types of visualizations, especially interactive visualizations, that are useful and successful. I found the section about "change blindness" to be fascinating. As Healey explains, a main character in a movie can change gender without the majority of the audience ever noticing. I think this is an important principle to understand when creating visualizations because it suggests that in certain scenarios viewers of a scene may be tricked, inadvertently, into missing detail about a particular image. This seems particularly applicable to Tufte's small-multiples, where changing one aspect of the design in a multiple may be completely missed or glossed over by the viewer.



[add comment]
Personal tools