Abstract:Building extraction is a fundamental task in remote sensing image processing and plays a crucial role in modern engineering. A number of studies perform building extraction by pixel-wise segmentation and have achieved impressive performance in producing binary (building and nonbuilding) segmentation masks. However, it is challenging to convert these segmentation masks into a set of vector polygons required for geographic and cartographic applications. To combat this issue, contour-based methods are proposed to directly predict a set of building polygons. However, the accuracy of their generated building polygons might be compromised as they overlook the geometric characteristics of buildings or situations where some building vertices are not predicted. To tackle these challenges, this article proposes an iterative polygon deformation algorithm (IPDA), which includes two essential modules: initial polygon generation and missing vertex recovery. The former generates a building polygon for each instance based on the geometry of buildings, while the latter iteratively recovers building vertices that have not been predicted. Experiments conducted on five challenging datasets show that IPDA achieves significant improvements while maintaining less inference time. Furthermore, the proposed IPDA can also be extended to other contour-based methods, enhancing their performance. The code is available at https://github.com/zhuyh1223/IPDA/.

Iterative Polygon Deformation for Building Extraction