WCSE 2023
ISBN: 978-981-18-7950-0 DOI: 10.18178/wcse.2023.06.019

Non-local Dense RepPoints for Instance Segmentation in Remote Sensing

Xinmin Xiang

Abstract—Recent advances in deep convolutional neural networks have significantly improved instance segmentation. In large-scale remote sensing images, however, the high density, arbitrary shapes and orientation, large aspect ratios, and huge scale variation of the objects pose significant challenges to general instance segmentation algorithms. In this paper, we propose a new framework, called non-local dense RepPoints, to improve the performance of instance segmentation in remote sensing images. First, we propose a hierarchical non-local block that iteratively integrates global information, and our method enables the model to accurately represent the relationship between two locations. Second, we enhance Dense RepPoints by designing an efficient dynamic vector to more efficiently model the objects by a large number of adaptive points. We conduct experiments on the iSAID dataset and compare our method with several commonly-used state-of-the-art networks. The experimental results demonstrate that our proposed approach can achieve promising results.

Index Terms—Remote Sensing, Instance Segmentation, Non-local, Dense Reppoints

Xinmin Xiang
College of Hydrology and Water Resources, Hohai University, CHINA


Cite: Xinmin Xiang, "Non-local Dense RepPoints for Instance Segmentation in Remote Sensing" Proceedings of 2023 the 13th International Workshop on Computer Science and Engineering (WCSE 2023), pp. 129-137, June 16-18, 2023.