I’ve been exploring the possibilities of historical face-averaging here for some time now—that is, “averaging” groups of facial images from successive periods and then arranging them into timelines so that we can compare them side by side or, better yet, watch historical trends unfolding before our eyes as video animations. The reason it’s been a year since I last blogged about this technique is not that I’ve lost interest in it, but rather that I’ve been slowly chipping away at a longer-term project along these lines: “The Fashionable Face.” It’s still far from complete, but I’d like to go ahead and share some of the results I’ve achieved so far by way of a progress report. (For my earlier forays into historical face-averaging, a.k.a. prosopochrony—which have involved yearbook photos, portraits on baseball cards, contestants and winners in the Miss America pageant, Fayum mummy portraits, and late medieval and early Renaissance painting—see the posts here and here from July 2014, and here from December 2014.)
My goal with “The Fashionable Face” is to use face-averaging to display changes in the “fashionable” female face over time.
For example, by averaging all the color images of women’s faces found in any leading fashion magazine during a given period (including the cover, articles, and advertisements), we should be able to generate a composite face that reveals the quintessential “look” of its era more accurately than any of the individual source images could by itself. Here are some averages I created from Vogue starting with a recent year (2014) and then skipping backwards from there decade by decade:
How representative are these averages of the fashions of their times? Well, they do look much as I would have expected based on the stereotypes I have in my mind’s eye for periods such as the mid-1980s or mid-1940s. Moreover, if we compare my Vogue averages with another set I created from Harper’s Bazaar, dated one year earlier in each case, we can see that the averages for adjacent years are very similar to one another, and that they resemble each other more than they do any of the other averages.
The number of source images I’ve been using varies from case to case. For the examples running from left to right above, I respectively averaged 90, 95, 95, 150, 216, and 160 images from Harper’s Bazaar and 109, 203, 77, 209, 154, and 241 from Vogue. I began in January each time, and I never went beyond June, but the span of months ranges from six in the earlier examples to three in the later ones, due partly to the way I found the issues bound together into volumes and partly to the number of suitable images per issue. (Vogue cut down rather starkly on its use of color images during the seventies for some reason, which accounts for the comparatively low figure of seventy-seven usable source images available from January through June 1974.)
In any event, it seems that each of these images really is coalescing on a quintessential “look” that would emerge more and more clearly if more and more data were added, and that differences are obvious between decades but far less noticeable between individual years. That latter point is important. Change was presumably gradual rather than abrupt, and by filling in intermediate phases in the sequence of averages, we ought to be able to illustrate the evolution of the “fashionable face” through nearly imperceptible gradations. WatchCut Video has created some popular “time-lapse” videos on the theme of “100 Years of Beauty,” starting with this one, but the time-lapse element in their case entails showing a modern-day subject as she’s being made up retrospectively in the styles of different years—1910, 1920, 1930, and so on—rather than transitioning gradually from one phase to the next. So where they present just two different appearances for the years 1940 and 1950—
I chose to start with Vogue and Harper’s Bazaar partly for the practical reason that bound volumes of these magazines are conveniently available to me at the Wells Library at Indiana University Bloomington. However, the local run of Harper’s Bazaar begins in the 1940s, and there don’t seem to be enough usable color images in Vogue to yield good results once we start pushing back into the mid-1930s. So I’ve been unsure for a while as to how best to extend the timeline of my project further backwards than that. Recently, I’ve turned to the periodicals digitized as part of the Media History Project, including large numbers of movie magazines, and these have been yielding promising results back into the 1910s (the images shown here are based on the data for entire years):
My hope is that the averages based on movie magazines as we approach 1940 will closely resemble the averages based on fashion magazines from the same period, which I’d take as evidence that the two data sets are reasonably interchangeable for my purposes—but we’ll see how it goes. If they’re significantly different, that would be interesting too. I’m not sure where to turn next, in order to push back further into the 1900s, 1890s, and beyond—maybe to sheet music covers or some common genre of advertising art. Suggestions are welcome, as is help in tracking down source material.
My ultimate plan is to combine all of these images into an animated video, which might look something like the animations I made a while back from composites of Miss America winners (explained here, but also available on YouTube), including this one based on twenty-five-year averages:
I’m not far enough along to produce an animation like that from my “Fashionable Face” data, and won’t be for quite a while yet, although I can at least share an animated version of the sequence running from 1940 to 1950 shown above as a strip of still images (please note that this is just a quick-and-dirty prototype and still needs considerable work):
Fortunately, I don’t need to wait on my own results to demonstrate this kind of animated display more fully, thanks to a recently-unveiled research project carried out by Shiry Ginosar, Kate Rakelly, Sarah Sachs, Brian Yin, and Alexei A. Efros entitled “A Century of Portraits: A Visual Historical Record of American High School Yearbooks.” They came up independently with the idea of averaging senior yearbook photos by year and using the same data to trace the rise of the “say cheese” smile, as covered previously on Griffonage-Dot-Com (see here). More importantly for present purposes, they assembled and analyzed a much larger source base than I had, totaling 37,921 images in all. They’ve published some nice still images of male and female faces based on this data, averaged by decade—one pair for the 1900s, one pair for the 1910s, and so on. Their results (this is just an excerpt of a larger composite)—
The original researchers don’t seem to have pursued the idea of animating their averaged images themselves. However, Aymann Ismail of Slate took a version of their data set averaged by year, rather than by decade, and produced a time-lapse video from it which may be viewed here. It differs from the animations I’ve been making myself—and sharing on this blog—mainly in that the individual images switch out slowly enough for the viewer to grasp each one distinctly in turn. By contrast, I’ve been creating animations with higher frame rates, trying to achieve illusions of seamless change over time, more like you’d associate with a full-fledged motion picture. As you can imagine, I’ve been curious to find out what my approach could do with this other yearbook data set. Rather than starting over from scratch with the original data set, which would have been prohibitively time-consuming, I’ve taken the frames of Ismail’s video as a starting point, hoping that my use of them is transformative enough that nobody will object. Although the original data set and Ismail’s video both extend back to 1905, some years are missing (1907, 1917, 1918, 1920, 1921, and 1924), so a continuous sequence of annual averages is actually available only for the years 1925-2013, and I’ve limited my sample animations to those years.
First, here’s an animation that displays each year’s individual average for only 0.07 of a second, rather than half a second as in Ismail’s original video. The accuracy of each frame is limited by the number of source images contained in the data set for that year, but the time base is as precise as yearbooks allow, with each frame corresponding exactly to a single year. The result “works” after a fashion, but it comes off as rather flickery, jitttery, and twitchy.
The next animation again advances at a rate of one year per frame, but this time each frame is based on the combined data for two successive years. The accuracy of each frame is higher than before, since it’s averaged from roughly double the number of source images, but the time base is less precise, since each frame now corresponds to a two-year range: 1925-26, then 1926-27, then 1927-28, and so on.
And finally, here (below) is an animation based on eight-year composites. The averages are the most accurate yet, with each one averaged from roughly eight times the amount of data as in the first example, but the time base is also the least precise, with each frame representing an eight-year period: 1925-32, then 1926-33, then 1927-34, and so on. It advances at a rate of one year per frame as before, but this time I cheated a little—every other frame is interpolated (i.e., tweened). For example, the first frame is based on the actual data for the years 1925-32, and the third frame on the actual data for the years 1927-34, but the second frame is simply the average of the first and third frames. The idea is that the average faces for 1926-33 would probably be pretty darn close to the averages of the averages for 1925-32 and 1927-34, so that we can fudge things in the interest of saving time and effort.I’ve retained the dates from Ismail’s video throughout, partly because they help us keep track of where we are at each point in the time sequence, but also because their blurriness is a good visual guide to the “blurring” of the time base. The final example displays the kind of smoothness I’d like to aim for with “The Fashionable Face,” but ideally with more precision in the time dimension.
A few different variables are in play here, which are worth spelling out in order to survey the terrain (feel free to skip my bulleted list, though, if you’re not interested in the technical minutiae):
- The quantity of source images on which each frame is based. The more source images, the more accurate the average, presuming they’re suitably “representative” of something.
- The display time window, or span of time represented by each frame in the final animation. In the yearbook examples shown above, this was respectively one year, two years, four years, and eight years.
- The raw time window, or span of time represented by averages created directly from the source data. So, for instance, I might first create a sequence of averages with a raw time window of half a year each, accumulating them over time as convenient, and then combine them into overlapping averages with a two-year display time window for the final animation. The raw time window in all the above yearbook examples was one year, even though the display time window varied from case to case. Different kinds of source material will sustain different minimum time windows. For instance, data from yearbooks can’t be organized practically into time windows of less than one year. Data from monthly magazines could conceivably be organized into monthly time windows, but the differences between individual months aren’t likely to be very meaningful for tracking something like fashion trends, so a longer time window might make sense in that case. On the other hand, the raw time window can be made shorter than is needed to yield enough data for acceptably representative averages, as long as the results are combined in turn into composites with a longer display time window.
- The display incremental time resolution, or difference between the time spans represented by successive frames in the final animation. In the yearbook examples shown above, this was consistently one year, even though we achieved this in one case by interpolating frames.
- The raw incremental time resolution, or difference between the successive time spans into which the original source data was divided for averaging. This was one year for the first three yearbook animation examples, but two years for the last example.
- The frame rate. In the above yearbook examples, each frame is set to display for 0.07 seconds, which comes out to roughly 14.3 frames per second.
- The frame data overlap. To create a “smooth” animation, it’s helpful for some of the source data to overlap from frame to frame. So, for instance, the display incremental time resolution could be a single month, but the time window might be a whole year. One frame would then combine all the data for (say) January through December 1940; the next would combine the data for February 1940 through January 1941; the next would combine the data for March 1940 through February 1941; and so forth. Each frame would contain data for twelve months, but the data for each pair of adjacent frames would differ by only one month and would overlap by eleven months. The yearbook animations above have respective frame data overlaps of zero years, one year, three years, and seven years.
- The quick-and-dirty prototype animation I shared above based on the “Fashionable Face” data I’ve analyzed so far for the period 1940-1950 has a display time window of three years, a display incremental time resolution of one year, a raw time window of half a year, an inconsistent raw incremental time resolution, and an inconsistent frame data overlap.
In closing, here’s a nice little challenge for you. Can you guess the years represented by the following four “mystery” averages?
Please venture your best educated guesses in the comments section below, limiting them to specific years rather than ranges of years or decades—think of it like trying to guess the number of beans in a jar. If and when I’ve received fifty different sets of guesses, I’ll post the correct answers, so be sure to invite all your friends to participate if you’re eager to know what they are. At that point I’ll also reveal the “winner” whose guesses are most consistently accurate, for whatever the bragging rights may be worth. Have fun!
PS. (July 2016): It’s been nearly half a year since I issued this challenge, and nobody has yet taken me up on it. Hmm….