Making a facial recognition system for Southeast Asia

Additionally, Grab’s facial recognition backend was expanded on a set of spoof versions of images in its database. This allows it to better pick out false face data, such as people showing photographs to their phones or wearing masks of other faces.

We relied on data annotators to manually label, tag, and classify data to synthesise image spoofing cues to train our models. Common spoofing cues include distortion and interferences that appear when cameras capture images of screen replays, or printed masks.

A database for Southeast Asian faces

Facial recognition systems are only as good as the data they are trained on.

According to a report published by US-based National Institute of Standards and Technology in 2019, most facial recognition algorithms exhibit bias. The 189 major facial recognition algorithm systems studied falsely identified Black and Asian faces 10 to 100 times more often than they did white people.

This is where Grab’s efforts to collect large and comprehensive face datasets come into play. Our database covers a wide range of demographic groups in Southeast Asia, and has allowed us to build a far less biased facial recognition model that serves the region better.

And as mask-wearing dominated everyday life in the past few years, we also wanted to develop a facial verification model that can verify identities even when people have their masks on. We began to add masked selfies into the training data, allowing the model to process both masked and unmasked selfies.