FlowVOS: Weakly-Supervised Visual Warping for Detail-Preserving and Temporally Consistent Single-Shot Video Object Segmentation

• DOI: 10.48550/arxiv.2111.10621
• Publication Date: 2021-01-01
• Authors: Julia Gong; F. Christopher Ho...; Serena Yeung
• Keywords: Image warping; Optical Flow; Dynamic Time Warping

We consider the task of semi-supervised video object segmentation (VOS). Our approach mitigates shortcomings in previous VOS work by addressing detail preservation and temporal consistency using visual warping. In contrast to prior that uses full optical flow, we introduce a new foreground-targeted warping learns flow fields from data. train module capture detailed motion between frames two weakly-supervised losses. object-focused foreground masks their positions target frame enables mask refinement with fast runtimes without extra supervision. It can also be integrated directly into state-of-the-art networks. On DAVIS17 YouTubeVOS benchmarks, outperform offline methods do not use data, as well many online Qualitatively, show our produces segmentations high consistency.