MIT develops 'gaze-tracking model' that can work out what people are looking at

Eye-gaze direction could be used in-store by retailers to learn what catches customers' attention

Researchers at MIT have proposed a new 'gaze-tracking model' that, they claim, is capable of detecting where people are looking in videos.

Past studies have suggested that in order to better understand the state of mind of an individual, one must be able to observe some specific behavioural cues. Eye-gaze direction is one such social cue.

In theory, 'gaze' is described as the act of seeing or being seen. It is a strong form of non-verbal communication and attention during social interactions. The social applications of gaze have been explained in detail by various philosophers in their writings.

Now, researchers at the Massachusetts Institute of Technology (MIT) have proposed a new gaze-estimation model, detailed in a paper published at this year's ICCV 19.

In their paper [PDF], the researchers talk about a large-scale gaze-tracking dataset, along with a method for robust 3D gaze estimation in unconstrained images.

The dataset, dubbed Gaze360, consists of labelled 3D gazes across a wide range of head poses and distances. The dataset uses 238 subjects in outdoor and indoor environments, and according to researchers, it is the largest 3D gaze dataset by variety and subject available in public domain.

The proposed 3D gaze estimation model extends the existing models to include temporal information and to directly output an estimate of gaze uncertainty.

Researchers carried out an ablation study to evaluate the model and its generalisation capabilities. They also performed a cross-dataset evaluation to assess the value of the novel Gaze360 dataset. In this cross-dataset evaluation, they used three other benchmark datasets, namely, RT-GENE, Columbia and MPIIFaceGaze.

The researchers found that the best results were achieved using their Gaze360 dataset as opposed to other three datasets. They also demonstrated an application of their model for estimating customer attention in a supermarket setting.

Last year, Microsoft also introduced an eye-control feature for Windows 10 to enable users to operate an onscreen mouse, keyboard, and text-to-speech experience using only their eyes. The feature is especially useful for people with disabilities, according to the software maker.