Towards light-compensated saliency prediction for omnidirectional images

Abstract

Omnidirectional or 360-degree images is becoming very popular in many applications and several challenges are raised because of the nature and the representation of the data. The saliency prediction for such a content opens the door to many problems linked to the geometric distortions, lighting variation, … In this paper, we propose a saliency model taking advantage of the large literature of 2D saliency and offering three major adjustments related to the nature of 360-degree images : 1) illumination normalization to account for the variability of lighting over the scene, 2) distortion compensation to handle the conversion problem from the sphere to the equi-rectangular representation, and 3) equator bias to incorporate the perceptual property according to which the human gaze is biased towards the equator line. The obtained results showed an improvement of the performance of the 2D saliency when using the above adjustments for omnidirectional images.