Towards Efficient Visual Attention Prediction for 360 Degree Videos

Visual attention prediction refers to the ability to predict the most visually important or attention-grabbing areas in a scene, and emphasize them to create an engaging and realistic experience for the user. These technologies require real-time processing of high-quality visual content to maintain user engagement and immersion. As such, it is necessary to use lightweight models that can predict the most important regions of a scene without incurring large computational cost. The contribution of this work is the development and evaluation of a lightweight model for visual attention prediction, which serves as a baseline on public datasets. We study various model design choices and their effects on the performance and efficiency. We also study the effect of a model compression technique, namely self-distillation.