ITEM: Forget facial recognition – the hot new AI-powered surveillance technology in China is emotion recognition. Or at least it is if you ask vendors exhibiting the technology at Chinese security industry expos.
Writing for the Financial Times, journalist Sue-Lin Wong attended the recent China Public Security Expo in Shenzhen, and her Twitter thread here is worth checking out just to get an idea of the scope of the surveillance technology industry – and how many domestic Chinese companies are playing in that field.
That includes tech giants like Alibaba, Tencent, Baidu and Huawei, who are also targeting the public security segment, mainly via their AI technologies that can power security solutions such as facial recognition, which is being deployed throughout China.
Emotion recognition is a subset of facial recognition that seeks to identify not just a person’s face, but determine their emotional state. According to the Financial Times, Chinese authorities are already rolling out the technology with the hopes of adding it to their arsenal of surveillance tools to spot criminals and even predict criminal behavior. Emotion recognition can purportedly do this by analyzing a suspect’s facial expression for signs of nervousness, aggressiveness, elevated stress levels and other indicators that they might be up to something.
Just how widespread the technology is in China seems to depend who you ask. A spokesperson from facial recognition startup Megvii said emotion recognition technology is being widely used within the Chinese government. But a spokesman for Baidu said “only a few schools and public security bureaus” use such technology right now.
One expo speaker said emotion recognition tech is being rolled out now at airports and train stations nationwide (mostly at customs checkpoints) to catch terrorists and smugglers – especially in Xinjiang, which already has already become a testing ground for cutting edge surveillance tech that’s mainly used to police its minority Uighur population.
However widely it’s being used in China, there’s certainly no doubt that Chinese authorities are interested in the possibilities – provided it works.
And therein lies the problem.
Emotion recognition has been in the pipeline for some time now. Big names like Amazon, IBM and Microsoft have been working on algorithms that claim to detect a person’s emotional state – not just in terms of crime prediction or public security, but for things like retail stores detecting when a customer is frustrated or annoyed, movie studios test-screening a new film, or HR staff determining suitable job candidates based on their attitude towards the interview.
The basic idea is simple enough – your facial expression conveys your emotional state. The underlying analogy is that a facial expression is a “fingerprint” that can be connected to the emotion producing that expression. For example, Amazon says its ‘Rekognition’ software can classify eight categories of emotion: happy, sad, angry, surprised, disgusted, calm, confused and fear.
Sounds simple – but it’s not. The facial expression recognition (FER) part alone is challenging. A recent study published in the academic journal Sensors notes that while FER systems can achieve accuracy rates of up to 97% in the lab, that rate drops to around 50% in real-world apps. Things like variances in lighting and the position of the head can throw off the system’s performance.
The bigger challenge is that emotional expression is a lot more complex than people (and possibly tech companies) realize. A recent academic study commissioned by the Association for Psychological Science and conducted over two years found that emotions are expressed in a wide variety of ways that don’t always correspond to facial expressions. Easy examples are situations where people cry with joy or laugh when they’re angry.
The study’s results indicate that there’s so much variety and nuance to navigate that emotion algorithms have to do a lot more than determine whether you’re smiling, scowling, laughing or whatever.
To be sure, many emotion recognition algorithms combine facial expressions with other metrics (tone of voice and eye movement, for example) to determine the corresponding emotion. In fact, there’s a lot of activity in developing AI algorithms to detect emotion in people’s voices. (There’s also an arms race to defeat such algorithms.)
Either way, the upshot is it’s still early days for emotion recognition. And all this is before we get into issues like racial bias in AI and how facial recognition systems don’t work well on non-white faces.
Presumably the technology will improve over time. The question is whether it’s ready now for use in public security apps – in China or anywhere else – and more importantly, what will be the consequences for people when the algorithm gets it wrong.