2D-Based Saliency Prediction Framework for Omnidirectional–360° Video

This paper investigates the performance of 2D-Based visual saliency models applied to omnidirectional videos. Indeed, most existing 2D-Based saliency prediction techniques for dynamic and static scene analysis consider essentially the planar projection format. Moreover, some of the well-known prediction saliency models in the literature were implemented and widely tested with specific equirectangular (ERP) and cubemap (CMP) projections in order to track visual attention in 360° videos, using head-mounted displays (HMDs). Furthermore, the geometric difference between the two spaces raises a critical bound, on the performance of 2D models adapted to 360° data using some ERP and CMP projection methods. The main issue is the merging of 2D saliency probability distributions of some ERP and CMP projections to deal with the saliency prediction in 360° videos. Our framework can integrate new 2D saliency models without requiring any adaptation of the existing prediction algorithm. Our experiments on Salient-360 data sets show that the proposed framework integrating the 2D-based models in Omnidirectional saliency prediction is competitive with the state-of-the-art, at the head and eye movement saliency prediction.