Video Saliency Prediction Based on Spatial-Temporal Two-Stream Network

In this paper, we propose a novel two-stream neural network for video saliency prediction. Unlike some traditional methods based on hand-crafted feature extraction and integration, our proposed method automatically learns saliency related spatiotemporal features from human fixations without any pre-processing, post-processing, or manual tuning. Video frames are routed through the spatial stream network to compute static or color saliency maps for each of them. And a new two-stage temporal stream network is proposed, which is composed of a pre-trained 2D-CNN model (SF-Net) to extract saliency related features and a shallow 3D-CNN model (Te-Net) to process these features, for temporal or dynamic saliency maps. It can reduce the requirement of video gaze data, improve training efficiency, and achieve high performance. A fusion network is adopted to combine the outputs of both streams and generate the final saliency maps. Besides, a convolutional Gaussian priors (CGP) layer is proposed to learn the bias phenomenon in viewing behavior to improve the performance of the video saliency prediction. The proposed method is compared with state-of-the-art saliency models on two public video saliency benchmark datasets. The results demonstrate that our model can achieve advanced performance on video saliency prediction.