French researchers have developed virtual characters that meet a person's gaze just like another human. The software allows the characters to look at scenes and people the way humans do.
Scientists say the goal is to make virtual humans and perhaps humanoid robots easier to relate to. The software developed by Gérard Bailly and colleagues in the GIPSA Lab at the Institut National Polytechnique de Grenoble, France, mimics human gaze patterns.
Their characters are capable of saccades - the unconscious jerks which constantly makes our eyes dart around - tracking moving objects like humans, and fixing their gaze on the same features as humans for similar periods.
The scientists say the new software is based on a pioneering model devised in 2003 by Laurent Itti and others at the University of Southern California, Los Angeles, US, to mimic human vision.
According to the New Scientist magazine, the model deals with scenes in three ways: looking for 'saliency' or the most visually outstanding parts in a scene, 'pertinence' or the most important parts, and 'attention', which temporarily inhibits regions that are no longer interesting.
Bailly's team, however, added several extra mechanisms; an 'attention stack' that tries to better mimic the way humans rank interesting areas, while another module that recognises certain familiar objects, such as faces.
These allow the software to focus a character's eyes on particular scene details at relevant times, like the eyes and mouth on a face when communicating, for example.
The team made a third addition to Itti's model, as well, a 'retinal filter' that simulates the difference between peripheral vision and the high-resolution information gathered by the centre of the retina. Bailly said the team has already tested their model in face-to-face trials where people interacted with a humanoid using the software.
"We found that the robot's gaze patterns were comparable to the ones recorded on human subjects observing the same scene," said Bailly. He said the new attention model would be crucial for giving virtual characters human-like movements and good social skills. It would also help them act human because, by using the same vision strategy as a human, they will gather the same information a human person would from a scene, Bailly added.
"These agents should be able to analyse the scene they're they are interacting with," said Bailly. "This research is important because it focuses on adding a social aspect [to characters]," added Christopher Peters of the University of Paris VIII, who has researched similar problems.
"But it also raises difficult questions and challenges -- such as how to model competition for visual attention between different stimuli, like multiple faces and conflicting emotions," Peters added.