Multi-stream Temporally Enhanced Network for Video Salient Object Detection

Video salient object detection (VSOD) aims at locating the most attractive objects in a video by exploring the temporal and spatial features. Despite the progress made in existing VSOD models, there is still a tendency to prioritize temporal features over spatial features. This bias has been shown to result in the imprecise localization of salient objects in the spatial domain and the accumulation of interframe noise in the temporal domain. We aim to investigate saliency cues collaboration in the spatial domain, along with a straightforward, yet efficient approach for temporal feature extraction. This paper proposes a multi-stream temporal enhanced network (MSTENet) for VSOD. The MSTENet employs a multi-stream structure that leverages both foreground and background supervision to extract collaborative saliency cues in the spatial domain. Meanwhile, the temporally enhanced module focuses on extracting motion information by enhancing distinctive regions in the current frame relative to adjacent frames. The proposed framework seamlessly integrates these two components to achieve more accurate and efficient VSOD performance. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art performance on five benchmark datasets.