End-to-end Animal Matting

Abstract

Image matting is a widely studied low-level vision problem that aims to provide a detailed foreground alpha matte from a single image, benefiting a wide range of downstream applications. However, most of the prevalent matting models are requiring extra manual intervention such as trimap or scribble. Besides, the lack of large-scale real-world annotated data has also caused poor generalizability in learned deep models. In this paper, we propose a novel end-to-end matting method called GFM along with a real-world, high-quality, category-wised animal matting dataset called AM-2k to address the above issues. The proposed end-to-end matting model GFM is short for Glance and Focus Matting Network, aims to conduct simultaneously trimap generation and matting by sharing one encoder and going through different decoders in separate branches. The design of GFM can help extract local and global information within one stage training process. Without the need for any extra input, GFM surpasses the previous state-of-the-art in performance on all evaluation metrics. Our proposed AM-2k consists of 20 categories mammals animals and 200 high-quality image for each category. We manually generate accurate mattes for each of them. Based on this dataset, we also set up three evaluation tracks, MIX-Track, DA-Track and CW-Track which can benefit the research on end-to-end matting, trimap-based matting, domain adaptation for matting and few shot learning. Extensive experiments and comprehensive analysis are performed on the AM-2k dataset to validate the effectiveness of GFM and its superiority over representative state-of-the-art methods. Various visual results can be found in Chapter 4 and Appendix.