How to handle large archives challenges: understanding and visualizing trends

Maternity wear, 1950 to 2011 in Australian women magazines

Summary
The project
The data set
Trend identification
What happened to tartan?
Possible futures

What happened to maternity wear during the last 6 decades?
Are there recurring features?
What fashion trends can and Artificial Intelligence identify and trace in a mass of photos?

A lot, yes, plenty: we were able to map trends in styles, fabrics & patterns, spotting recurrences and persistence of traits and possible cycles. It is time for tartan to come back?
What we are about to describe is a special educational project that led us to a better understanding of our computer vision tool and offered substantial insights on how to build on company heritage.

The project

The Karl Hubenthal Morgue Of Research Imagery donated by Boob Staake to Columbia University’s Comics and Cartoons Library.

Companies have histories, every aspect of modern life and industry has been recorded or photographed, so much so that archives have become a cornerstone of brand images and repositories of values, along with a reservoir of solutions to past challenges.
The wealth of information that resides in the textual and visual documents are capable of shedding light on the present and future adventures when read through a lens that highlights trends, variation, interruption, continuities, think of fashion brands, magazines and brands.
We might say we’re delving into the Culturomics field, applying our technology to image repositories instead of texts (on this occasion) to trace cultural, social and industrial currents and trends.

In the case at hand, we worked together with Anna Anisimova, researcher and PhD candidate at the Royal Melbourne Institute of Technology, to help in a research project about fashion for pregnant women. Anisimova’s project aims at describing trends in maternity wear, the correlation of its features, providing social and cultural interpretations of changes over time. In order to do so she elected to use women’s periodicals as the main source of information for the research. Describing, analyzing and interpreting images collected from two fashion and lifestyle periodicals, Australian Women’s Weekly and Vogue Australia, is where Nextatlas comes in.

The data set

This slideshow requires JavaScript.

For the purpose of this project, we have been working with an archive of 1,035 images (retrieved from Australian on-line National archives), arranged by date, from which to extract information applying the Clarifai/Nextatlas joint style models.
Nextatlas has been working for a longtime with Clarifai technology: computer vision and results analysis have proven very successful and useful in the past.
To be clear, the style model is a neural network. It learned the various styles by studying tens of thousands of clothing-style images, to form an understanding of which patterns are important for the various styles. We used two separate models for styles and fabrics & patterns.
Our data scientists were faced with multiple challenges. Images had been scanned from the magazines pages, resolution was not homogeneous varying from 144dpi to 300dpi, AM and FM screening patterns posed another problem: printing quality has come a long way from the 1950s to today, another thing to keep in mind when dealing with such source materials.
An average of 16 pics per year with files sizes ranging from 48KB to 681KB, from 600x800px to 186x800px in dimension.
A wide variety when limiting consideration to data information, but to a human eye the repository offers even more details: it includes magazine covers, advertisement pages, with a whole array of superimposed writing in various fonts and sizes, artists drawings and sketches, staged photos and more gossip style ones.

Trend identification

The tuned computer vision engine scored each picture file against a set of styles (“90s now”, “Boho”, “Dapper”, “Edgy”, “Minimal”, “Remix”, “Workwear chic”) and some fabrics and patterns (“Brocade”, “Camo”, “Denim”, “Eyelet and lace”, “Gingham”, “Polka Dot”, “Stripes”, “Tartan”).

Let us show you an example and elaborate on it some more.

But first a striking spot-on image
Diana spencer koala sweaterThis photo of a pregnant princess Diana was taken in 1982 at a polo match and features the now famous bright blue ‘Blinky’ koala sweater.
It is a rather good example of what the system identifies as “Remix”.

A similar image appeared on the June 1982 issue of Australian Women’s Weekly magazine, of which we had a lower resolution version with some digital artifact noise and it scored 55% as “Remix”, 33% “90s now” and close to zero in all other styles.

 

It is important to note that the style model was trained mainly through contemporary imagery, hence, its accuracy is more dependable with recent years, higher resolution photos. Here are a couple of photos with a score close to 100 to exemplify at best the “Remix” style.

 

A part from single stand out images our intent, together with Anna Anisimova, was to identify trends arranged over timelines for the aforementioned styles and fabrics & patterns. We have run our models for each of the digital image we were provided with, gathering a score over 100 for every image in all traits.

The results were fascinating, and will be fully available in Anisimova’s final research, facing us with timeline curves visually exemplifying fashion trends.
Boho”, for example, had peaks at the end of the 60s and 70s, again in the mid 90s and again around 2010, its features clearly correlate with social and political movements (from anti-establishment to hipster).

What happened to tartan?

What happened to “Tartan” across the seven decades? We don’t see it heavily featured anymore but when was it common and would it be an interesting style to bring back?
To give an example of the complete set of data each image gave origin to, this photo appeared on the April 1972 issue of Australian Women’s Weekly.

Tartan 99,97416
Denim 97,68634
Gingham 0,31137278
Stripes 0,08627981
Camo 0,00101266
Eyelet and Lace 0
Leather 0
Polka Dot 0
Brocade 0
Eyelet and Lace 0

 

And this is the downward curve that describes the almost extinction of tartan in maternity wear in the sources chosen for the research.
Tartan timeline
The mid 50s show the highest peak in tartan presence for our repository, then a first steep decline followed by multiple bumps and further declines bringing to a flatline after the mid 90s. The examples provided on the timeline also highlight a transformation of the typical tartan pattern into a less recognizable version, farther away from Scottish tradition.

Possible futures

The possible applications of this model analysis are multiple and fascinating, given a larger sample and custom style models as deemed relevant by market or industry or cultural movement, the outcomes would concur to the creation of sector overviews and offer fertile ground to novel and creative market insights.
This research opens a variety of possibilities to build strategies on top of brand heritage: visual history has overcome written texts in archives and offers itself as an identity storage waiting to be used and better understood. As stressed in the Culturomics research every aspect of recorded history can become instrumental in developing future strategies: are we repeating ourselves? Is there a possible continuity? What were the consumers’ answers to our past choices?
And many other questions we’ll now be able to answer swiftly.