F2Net: Learning to Focus on the Foreground for Unsupervised Video Object Segmentation

Although deep learning based methods have achieved great progress in unsupervised video object segmentation, difficult scenarios (e.g., visual similarity, occlusions, and appearance changing) are still no well-handled. To alleviate these issues, we propose a novel Focus on Foreground Network (F2Net), which delves into the intra-inter frame details for foreground objects thus effectively improve segmentation performance. Specifically, our proposed network consists of three main parts: Siamese Encoder Module, Center Guiding Appearance Diffusion Dynamic Information Fusion Module. Firstly, take siamese encoder to extract feature representations paired frames (reference current frame). Then, Module is designed capture inter-frame (dense correspondences between reference frame), intra-frame original semantic frame. Different from Anchor Network, establish Prediction Branch predict center location leverage point information as spatial guidance prior enhance extraction, representation considerably focus objects. Finally, automatically select relatively important features through aforementioned different level features. Extensive experiments DAVIS, Youtube-object, FBMS datasets show that F2Net achieves state-of-the-art performance with significant improvement.