Saliency Prediction Network for $360^\circ$ Videos

Abstract

Panoramic videos are becoming more and more easily obtained for common users. Although these videos have 360° field of view, they are usually displayed with perspective views, which needs the saliency informations for viewing angle selection. In this paper, we propose a saliency prediction network for 360° videos. Our network takes video frames and optical flows in cube map format as input, thus it does not suffer from image distorations of panoramic frames. The network is composed of feature encoding module and saliency prediction module. The feature encoding module extracts spatial and temporal features. Then these features are processed by a decoder and bidirectional convolutional LSTM for saliency prediction. To more thoroughly mine the motion information, the temporal stream of feature encoding module accepts optical flows before and after current frame. We also incorporate the global feature of video frames, residual attention and Gaussian priors into the network by considering the viewing behavior of 360° videos, which is useful for performance improvement. To evaluate the performance of our method, we compare it with three state-of-the-art saliency prediction algorithms on two publicly available datasets. The experimental result has shown the effectiveness of our method, which gets the best performance.