Geometric feature based approach for 360° image saliency estimation

This paper presents a 360° image saliency estimation technique. Visual perception of a scene is highly motivated by its geometrical features. The proposed approach combines standard low-level features along with multiple geometry-based features for saliency estimation. Different from the existing saliency approaches, here we combine features such as shape, ridge, corner, and orientation that reflects the geometry of an image. The proposed model is divided into two modules. The first module subsamples input image and generates activation maps for each feature. Subsequently, the second module refines the map by addressing equator biasness, non-uniform illumination and distortion correction due to 360° image projection. Experiments are performed on standard 360° image dataset and the obtained results show the proposed model outperforms existing saliency models.