Subjective Evaluation of Egocentric Human Segmentation for Mixed Reality

PubDate: December 2021

Teams: Nokia Bell Labs

Writers: Ester González-Sosa; Pablo Perez-Garcia; Diego Gonzalez-Morin; Alvaro Villegas

PDF: Subjective Evaluation of Egocentric Human Segmentation for Mixed Reality

Abstract

Augmented Virtuality (AV) represents the combination of an immersive environment and a selection of the egocentric local reality. The latter is carried out by applying a particular image segmentation method selected according to the specific application requirements. These image segmentation algorithms can be evaluated from a computer vision perspective using traditional quantitative metrics such as Intersection over Union (1oU), Pixel Accuracy (PA), among others. As AV applications are designed to create experiences for real users, it is also desired to bare in mind their subjective feedback. Thus, in this work we compare segmentation methods from a perceptual point of view, by conducting subjective experiments with a group of users. Subsequently, we compare obtained perceptual results with algorithmic metrics, showing a significant correlation between them. However, we also found subtle differences that could influence decision-making.