Bridging Composite and Real: Towards End-to-End Deep Image Matting

Extracting accurate foregrounds from natural images benefits many downstream applications such as film production and augmented reality. However, the furry characteristics various appearance of foregrounds, e.g., animal portrait, challenge existing matting methods, which usually require extra user inputs trimap or scribbles. To resolve these problems, we study distinct roles semantics details for image decompose task into two parallel sub-tasks: high-level semantic segmentation low-level matting. Specifically, propose a novel Glance Focus Matting network (GFM), employs shared encoder separate decoders to learn both tasks in collaborative manner end-to-end Besides, due limitation available task, previous methods typically adopt composite training evaluation, result limited generalization ability on real-world images. In this paper, investigate domain gap issue between systematically by conducting comprehensive analyses discrepancies foreground background We find that carefully designed composition route RSSN aims reduce can lead better model with remarkable ability. Furthermore, provide benchmark containing 2,000 high-resolution 10,000 portrait along their manually labeled alpha mattes serve test bed evaluating model’s Comprehensive empirical studies have demonstrated GFM outperforms state-of-the-art effectively reduces error. The code datasets will be released at https://github.com/JizhiziLi/GFM .