There are three leaders of countries in the news quite a lot at the moment. And each appears to have a set of expressions when in the public eye. So what can computer vision tell us about them?
Computer vision and facial recognition technology are both improving and becoming more accessible for anyone who wants to play around with face and emotion detection. So what do the algorithms make of country leaders, people who are in the public eye so often they tend to default to a familiar set of expressions for just about any scenario, no matter how inappropriate.
The three leaders are: Donald Trump, president of the United States of America; Vladimir Putin, president of Russia; and, Kim Jong-un, supreme leader of the Democratic People’s Republic of Korea, aka North Korea.
My initial hypothesis was: Trump tends to be very expressive; Putin looks bored a lot; Jong-un smiles a lot, but is it a real smile?
I did a highly scientific search for images, by googling their surname and scanning through the first set of images that appeared. For each subject, I picked 5 images and tried to balance between their more familiar pose and its opposite. The joy of copyright means I can’t embed the images here. Instead, I’ve created rough sketches of the pictures. If you click on a sketch, it will display the source photo in a separate tab/window in the browser.
First challenge to the hypothesis – there were a lot more pictures of a smiley-Putin than I anticipated… Unsurprisingly, it was much harder to pick photos of Kim Jong-un in a natural or off-guard pose than for Trump or Putin.
For the analysis, I used Microsoft’s Cognitive Services, specifically the Computer Vision API and Emotion API. Computer Vision gives an estimate of the gender and age of the face detected, Emotion gives a score of the emotions detected, using the standard range of emotions: happiness, sadness, surprise, anger, fear, contempt and disgust. With neutral to make up the balance. The total score is 1 with a value assigned to each emotion detected.
Some vital statistics to be aware of when evaluating the face detection results (source: Wikipedia):
|Trump||male||71 (b. 1946)|
|Putin||male||64 (b. 1952)|
|Jong-un||male||33 (b. 1984)|
It should be noted that I didn’t check how old any of the photos are, although the goal was to select recent ones from news articles. But there may be some error in matching the current age of the person to his age when the photo was taken. Also, fairly low resolution images were used, which affects the accuracy of the algorithm. It would be expected to have quite a wide range in age estimates.
Here are the results. (Reminder: click on an image to view the source photo analysed)
Trump’s emotions were the most mixed which is fitting, given he is definitely the most expressive of the three. And, apart from one of Jong-un’s photos, Trump was the only one to register significant sadness, anger, disgust and contempt in any of the photos. His gender was not in any doubt and age was mostly under his actual age. One jumped higher, no.3, but the forced grin was not flattering.
Putin’s emotional range was much more restrained compared to Trump, and switched between two modes – neutral and happy. Whatever he might actually be thinking, his public persona appears to be kept well under control. I thought the algorithm would pick up a bit of anger, disgust or contempt. But not really… Putin’s expression, to a computer at least, is neutral when not displaying playful expressions. And, goodness, he churns out more of those than I expected.
Again, the algorithm had no problem deciding it was a male in the picture. And apparently a well-maintained one. Putin’s visual age came in well under his actual age.
Jong-un is similar to Putin in displaying a very limited range of emotion. And he is the only one to score the magic 1.0 for a single emotion – happiness. Whether that means he’s actually happy is for others to guess… But he’s certainly got a staple smile that works for the algorithms. Skimming through the photos in search results, in the large majority (that were authentic – there are a lot of photoshopped edits…) Jong-un’s expression suggests an awareness of cameras present. The only image that looks off-guard is the 4th one and, interestingly, that is the only one that picked up a range of emotions including registering disgust.
And again, the algorithm was happy to confirm the gender as male in all pictures. Age estimates were not bad apart from one, with most coming just above or below his actual age. The 3rd picture is an anomaly and I suspect a higher-resolution picture would reduce that age estimate.
Well of the three personalities, two seem to maintain a strong awareness of cameras always being present and show restraint in the display of visible emotions. No surprises for the one who reads more like an open book… but maybe that’s what maintains his popularity. Is there growing fatigue with manufactured and tightly-controlled personalities?
In terms of the algorithm, there are similar shortcomings to tests run last year. Sideways shots are still harder for the machines to figure out. They prefer full frontal shots… And I have unscientifically decided that the algorithm is indeed sexist as well as racist. Presumably because the samples used to build/improve the algorithm are going to feature more women in make-up, the algorithm tends to add years to a woman’s age if you aren’t fully moisturised and prepped for the photo. Men get off lightly 🙂
For those interested in the details, the following images show the results as they were produced in Python.