The Research Samples series recounts the experiences of members of the student research community. They share a glimpse into their graduate projects.
Justine Giroux, a doctoral student in the Faculty of Science and Engineering, is interested in the human perception of virtual lighting, which is used in a number of artistic fields. It was her interest in science and her creative side that motivated her to undertake a research project in the laboratory of Jean-François Lalonde, Assistant Scientific Director at the Intelligence and Data Institute.
The research group has been working for several years on methods for extrapolating the lighting conditions of an image and recreating them on a virtual object. "The brain can be easily tricked on certain aspects like lighting, but it’s good at perceiving what seems unrealistic," says the PhD student.
The aim of his project is to create a metric for discovering the most realistic lighting conditions of a virtual object according to the human eye. "Current metrics are not suitable. They compare light and its intensity pixel by pixel, but what they consider adequate doesn’t correspond to human perception," explains Justine Giroux.
To develop a new metric that addresses these shortcomings, the PhD student needed to know what human perception was. She therefore created an experiment to determine which lighting conditions were considered most realistic by the participants. Installed in front of a computer screen, in a completely dark room, they had to select their preferred lighting from several representations of a virtual object.
"There had to be no other light sources in the room so as not to influence the choice of image. The screen had to be calibrated, since its tint and lighting can influence perception. On two different screens, we would have had very variable results", stresses Justine Giroux, mentioning the challenges encountered.
After the human, it was the machine’s turn. Unsurprisingly for the doctoral student, none of the 15 metrics tested chose the same virtual representation as the participants.
With the data collected during the experiment, Justine Giroux was able to train an artificial intelligence model. She showed that a combination of existing metrics more accurately represented human preferences.
During her PhD, the young researcher will improve the metric using deep learning so that it can one day be used by the lighting estimation community. "The idea is to make work easier and more efficient. It has applications in many fields, such as virtual and augmented reality. It can also be used to help artists make their creations more realistic. We can also think of the advertising field", mentions the doctoral student.
The research results have been accepted as a paper at the Conference on Computer Vision and Pattern Recognition. Justine Giroux is delighted. "It’s the most important computer vision conference! Its selection criteria are very rigorous."
