AI taught to determine the time by the hands of the clock

Reading the time from analog clocks is surprisingly difficult for computers, but now artificial intelligence can accurately handle this task. The AI, trained using images of clocks taken from different angles, was able to pinpoint the time the devices were showing.

Computer vision has long been able to read the time from a digital clock just by looking at the numbers on the screen. But analog clocks are much more complex due to a number of factors, including different designs and the fact that hands hide shadows and reflections.

Researchers at the University of Oxford have developed a system that can read the dial of an analog watch with an accuracy of 74% to 84%. Scientists trained a computer vision model on three sets of generated clock images from different angles.

These images, as well as the correct time shown by the clock, were used to train a neural network with an STN module that could distort an angled image to view it at exactly 90 degrees. Earlier versions of such neural networks, without the STN module, made mistakes when transforming images.

The creation of additional data sets helps the authors of a new development to teach the algorithm to correctly perform its task. The model was also trained on a set of videos with accelerating hand movements across the dial. Scientists "fed" the algorithm 4472 clock images.

When a model fails the task, it is most often due to the fact that it confuses the minute and hour hands of the same length. People can deal with this complexity by evaluating the different speeds of the arrows, but this obviously doesn't work with static images.

A source: arxiv, newscientist