An Improved BoxInst Model for Plane Instance Segmentation in Remote Sensing Images
Shangzheng Jiang, Qingzhong Jia, Fengkun Luo, Tao Yang · 2021
When traditional instance segmentation model is applied to the plane target in the remote sensing image, high-frequency co-occurrence background is easy to appear, which affects the segmentation precision. To solve this problem, this paper proposes an improved BoxInst model for the plane segmentation in remote sensing images. The improved model suppress high-frequency co-occurrence background by introducing background constraint loss. Foreground constraint loss is proposed to avoid trivial solutions caused by the background constraint loss. The improved model is verified on the plane data set, and the average precision of the proposed model is improved by 6% compared with the original BoxInst model.