Video Object Segmentation Using Multi-Scale Attention-Based Siamese Network

Video target segmentation is a fundamental problem in computer vision that aims to segment targets from a background by learning their appearance information and movement information. In this study, a video target segmentation network based on the Siamese structure was proposed. This network has two inputs: the current video frame, used as the main input, and the adjacent frame, used as the auxiliary input. The processing modules for the inputs use the same structure, optimization strategy, and encoder weights. The input is encoded to obtain features with different resolutions, from which good target appearance features can be obtained. After processing using the encoding layer, the motion features of the target are learned using a multi-scale feature fusion decoder based on an attention mechanism. The final predicted segmentation results were calculated from a layer of decoded features. The video object segmentation framework proposed in this study achieved optimal results on CDNet2014 and FBMS-3D, with scores of 78.36 and 86.71, respectively. It outperformed the second-ranked method by 4.3 on the CDNet2014 dataset and by 0.77 on the FBMS-3D dataset. Suboptimal results were achieved on the video primary target segmentation datasets SegTrackV2 and DAVIS2016, with scores of 60.57 and 81.08, respectively.