Reading the time on an analogue clock is surprisingly difficult for computers, but artificial intelligence can now do so accurately using a method that had previously proved tricky to deploy.
Computer vision has long been able to read the time from digital clocks by simply looking at the numbers on the screen. But analogue clocks are much more challenging because of factors including variation in their design and the way shadows and reflections can obscure the hands.
Researchers at the University of Oxford have developed a system that can read an analogue clock face, achieving 74 to 84 per cent accuracy on three sets of test images. They did so by training a computer vision model on computer-generated images of clocks as seen from different angles.
Advertisement
These images, along with the correct time shown by the clocks, were used to train a spatial transformer network (STN), which can warp an image taken at an angle in order to look at it face on. STNs have rarely worked when tested on photos of clocks from different angles before, because they warp the images incorrectly 鈥� but the synthetic data here helped it do this correctly.
The model was also trained on a set of time-lapse videos of clock faces, which improved accuracy when tested against 4472 images of clocks. When the model failed, it did so most commonly because it confused minute and hour hands that were a similar length. People can easily check in this situation, by looking at whether the hour hand鈥檚 position between numbers matches the time shown by the minute hand.
鈥淚t鈥檚 an extremely creative application paper,鈥� says at Google AI. 鈥淕iven the challenge with getting manually labelled data, the authors creatively use synthetic data and the fact that time progresses to automatically correct pseudo labels.鈥�
While reading clocks might seem a fairly niche task, Nagrani believes the underlying concept can be used with any type of analogue machine, including other scientific instruments.
Reference:
