DS-Net: Dynamic spatiotemporal network for video salient object detection

As moving objects always draw more attention of human eyes, the temporal motion information is exploited complementarily with spatial to detect salient in videos. Although efficient tools such as optical flow have been proposed extract information, it often encounters difficulties when used for saliency detection due movement camera or partial objects. In this paper, we investigate complementary roles and propose a novel dynamic spatiotemporal network (DS-Net) effective fusion information. We construct symmetric two-bypass explicitly features. A weight generator (DWG) designed automatically learn reliability corresponding branch. And top-down cross attentive aggregation (CAA) procedure facilitate Finally, features are modified by guidance coarse map then go through decoder part final map. Experimental results on five benchmarks VOS, DAVIS, FBMS, SegV2, ViSal demonstrate that method achieves superior performance than state-of-the-art algorithms. The source code available at https://github.com/TJUMMG/DS-Net.