PSNet: Parallel Symmetric Network for Video Salient Object Detection

For the video salient object detection (VSOD) task, how to excavate information from appearance modality and motion has always been a topic of great concern. The two-stream structure, including an RGB stream optical flow stream, widely used as typical pipeline for VSOD tasks, but existing methods usually only use features unidirectionally guide or adaptively blindly fuse two features. However, these underperform in diverse scenarios due uncomprehensive unspecific learning schemes. In this paper, following more secure modeling philosophy, we deeply investigate importance comprehensive way propose network with up down parallel symmetry, named PSNet. Two branches different dominant modalities are set achieve complete saliency decoding cooperation Gather Diffusion Reinforcement (GDR) module Cross-modality Refinement Complement (CRC) module. Finally, Importance Perception Fusion (IPF) according their scenarios. Experiments on four dataset benchmarks demonstrate that our method achieves desirable competitive performance.