Asymmetric Attention Fusion for Unsupervised Video Object Segmentation

In this paper, we introduce a novel Asymmetric Attention Fusion Network (AAF-Net) based on an attention mechanism to complete unsupervised video object segmentation. Firstly, we propose an asymmetric attention fusion module (AAFM) to aggregate the two source inputs to exploit the complementary representations between optical flow and RGB images. The asymmetric attention structure of AAFM is equipped to enhance feature information. Next, we design a feature correction module (FCM) to balance the information ratio between motion features and appearance features. The results of experimental evaluation obtained on several well-known benchmarking datasets, including DAVIS16, FBMS, and SegTrack-V2, deliver outstanding performance compared to the other segmentation networks based on optical flow, reflecting the merit and advantage of the proposed approach.