Screening and Multiscale Fusion Networks for Video Salient Object Detection
Zhu Huang, Jun Wang, Yandong Hou, Miaohui Zhang, Xing Ren · 2023
Optical flow images are widely used in video salient object detection as a tool to obtain temporal cues. How to effectively consider the fusion weight of spatio-temporal cues and ensure the accuracy of temporal cues in the detection process is the key to successfully detecting salient objects. Existing methods ignore possible errors in temporal cues and fuse all spatiotemporal cues with constant weights. In this paper, we develop a video salient object detection network that combines screening and multiscale fusion. The network is a two-stage network, with the first-stage screening network used to model against the stability of temporal cues, and the second-stage multi-scale fusion network used to explore the fusion effect under different spatiotemporal weights. Experimental results show that our method can achieve more reasonable spatio-temporal cue fusion states and validate the effectiveness of our final model on multiple datasets.