V-BMS360: A Video Extention to the BMS360 Image Saliency Model
Pierre R LeBreton, Stephan Fremerey, Alexander Raake · 2018
In this paper it is studied how existing image saliency models for head motion prediction designed for omnidirectional images can be applied to videos. A new model called V-BMS360 which extends the BMS360 model is presented. Due to the specific properties of omnidirectional videos, new key features had to be introduced: the first one is the introduction of a temporal prior that accounts for the delay needed by the users to explore the content during the first seconds of the videos. The second key novel idea is the consideration of camera motion and its consequences on the exploration of the visual scenes. It was translated into two key aspects new features: the “motion surroundness” and the “motion source”, and an informed pooling allowing to provide different strengths to different motion-based features based on a camera-motion analysis. All of these contribute to the performance of the new saliency model called V-BMS360.