期刊
EXPERT SYSTEMS WITH APPLICATIONS
卷 202, 期 -, 页码 -出版社
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.eswa.2022.117198
关键词
Instance segmentation; Macro-semantic morphological; Mask head; Long-range dependencies
类别
资金
- Fundamental Research Funds for the Central Universities [N2124006-1]
With the idea of divide and rule, there are two different forms of semantic features in two-stage instance segmentation paradigms: global features at the image level and instance features at the region-wise. The main distinction between these two macro-semantic morphological features lies in the relevance of neighborhood features caused by background noise. Therefore, we propose a more efficient methodology using Group-Inception and Asymmetric-Inception modules to enhance the representation capability of the network.
With the idea of divide and rule, there exist two different forms of semantic features flowing in the two stage instance segmentation paradigms. They are the global features at the image level and the instance features at the region-wise. The most significant distinction of the two macro-semantic morphological features lies in the different relevance of neighborhood features caused by background noise. Hence, we should consider different situations and make different schemes. Notice that the fields-of-view determines the range of local features that can be perceived in the convolution operation and implies the representation capability of the network. To this end, for FPN and Mask Head in two stage paradigms, we propose a more efficient methodology with Group-Inception and Asymmetric-Inception modules. This proposed methodology can act as a drop-in replacement to upgrade the plain convolution operation, which enables the network to look more via modeling long-range dependencies. Our method is simple yet effective. Quantitatively, we can significantly improve the state-of-the-art frameworks, including Mask R-CNN, Mask Scoring R-CNN, Cascade Mask R-CNN, and HTC by about 1.2%-2.2% AP on MS COCO test-dev yet with fewer parameters and FLOPs. Moreover, the proposed approach achieves competitive performances on the Scapes, KINS and SBD datasets. The source code of our method will be made available.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据