The invention of the camera in the early 19th century forever extended the boundaries of human vision and expanded the capabilities of the eye. Several decades later, the introduction of the kinetoscope and cinematograph gave rise to motion pictures and made the cameras the conduit for a whole new art form: cinema. Today, the camera not only mimics the eye, but helps thrill audiences, generate billions in box office dollars, and win shelves full of awards.
Awarding Films and Fashion
The Academy Awards have provided an annual opportunity to reward achievement in filmmaking for almost a century, but any Oscars enthusiast knows the fashion on display is as important as the films. Red carpet arrivals lead to Best Dressed and Worst Dressed lists which lead to memes and years of conversation.
Whether or not you’ve seen Dancer in the Dark, you’ve almost definitely seen the infamous “swan dress” that nominated musician Bjork wore to the Oscars in 2001 (more on that later).
To get in the Oscars spirit, Plainsight decided to shake up the conversation once again and introduce vision AI to Academy Awards. We dove into the Oscars fashion archives to develop a model for classifying suits and dresses. Our goal was to determine whether or not a vision AI model trained on Best and Worst-dressed lists would agree with popular consensus.
Classifying Oscars Fashion: Best and Worst
The annual awards season always seems to drag on forever, but developing our Oscars fashion model was quick and easy.
Collect the dataset
To create a model for classifying dresses and suits, we first needed to collect and classify a dataset for training. Fortunately, plenty of visual data is collected at the Oscars every year. Prior to the awards ceremony, hundreds of cameras line the red carpet to capture the arrivals of nominees, presenters, and other famous faces. It’s how outlets like The Hollywood Reporter and Entertainment Weekly can make their annual lists of the year’s best and worst-dressed stars. We collected images from the last decade worth of Oscars ceremonies, made note of whether journalists labeled them ‘Best’ or ‘Worst,’ and saved them in separate, appropriately labeled folders. This represented a rough approximation of consensus opinion on the last decade of Oscars fashion.
Upload the data and label it
Next, we uploaded the images to Plainsight and turned them into valuable training data by classifying each one with the platform’s labeling tool. The Labeler enabled us to designate each image either ‘Best’ or ‘Worst’ to help the model learn the difference. Since we were working with a relatively small dataset, manual annotation took less than an hour.
Train the model
Then, we created a model and put it to the test by training it on our labeled dataset. With just a few hundred images, Plainsight carried out the full training process in just a few minutes.
Review the results and continue testing
Finally, it was time to review our results and see how the model’s classifications compared to expert assessments of Oscars looks.
The Results: Did the Model Agree With the Experts?
Our model agreed with the fashion experts most of the time, but their occasional disagreements are potentially instructive. The results not only suggest that fashion is subjective, but reveal some of the inconsistencies in how Oscars fashion has been graded over the years.
Some of the model’s incorrect guesses make a lot of sense given the data it was trained with. Take Emma Stone’s 2012 dress by Giambattista Valli. At the time, the dress earned ‘best’ reviews and was classified as such for training purposes.
Our model nevertheless considered it an example of a ‘worst’ dress and made that assessment with considerable (79%) confidence.
How did the model get so confident in its incorrect guess?
Dresses like these two, worn by Anne Hathaway and Michelle Williams may have played a role. Note the superficial similarities between all three. They’re all red and ruffled, but only Stone’s was classified ‘best.’
Based on Hathaway and Williams’ examples, the model may have learned to associate red, ruffled dresses with the ‘worst’ classification. As such, it assigned Stone’s dress (which fits both criteria) the same classification as these others.
It should also be noted that these few dresses potentially had such an outsize impact on the model because of the fairly limited dataset. In a real-world use case, models would typically be trained with more visual data inputs. The Oscars only come once a year, so there’s only so much data to work with in this instance!
Testing the Model: Two Classic Oscars Looks
Just for fun, we put the model to the test on two classic Oscars looks. When eccentric Icelandic pop star Björk attended the Oscars in 2001, she did not collect the statue for Best Original Song. She did, however, enter the annals of Academy Awards history with “the swan dress.”
Gwyneth Paltrow was victorious in 1999, winning Best Actress in a Leading Role for her turn in Shakespeare in Love. While the film’s reputation has dipped, Paltrow’s pink Ralph Lauren dress has only seen its popularity improve. Though it earned mixed reviews at the time, the dress proved influential and remains an iconic example of Oscars fashion. Both dresses are among the very few garments with their own Wikipedia pages.
Our model agreed with contemporary assessments of both dresses, confidently assigning the ‘worst’ classification. Though the model may reflect certain biases, its guesses are totally unaffected by contemporary affection for the outfits and celebrities it sees.
Build and Train a Vision AI Model
Plainsight’s low-code vision AI platform is free for novices and experts alike to try. Register today to create and test models of your own for classifying fashion or helping address a range of everyday challenges.
Bennett Glace is a B2B content writer and cinephile from Philadelphia. He first watched the Academy Awards ceremony in 2005 and still considers them a guilty pleasure, even creating his own personal ballots each year.