<p>Sharp building segmentation is a challenging problem, due to the complex building shapes and extensive computational costs to process high-resolution and large-scale remote sensing images. To overcome the irregularity and redundancy of the building areas detected by the model, we present a novel and readily applicable gradient-preserving loss function, which utilizes the Sobel-Feldman operator to enforce constraints on building outlines. It leads to the generation of segmentation masks with straighter boundaries and sharper corners, which could avoid the need for complex polygonization algorithms in subsequent steps when applied to the problem of accurate building extraction. Our method achieves average precision values exceeding 97% on multiple datasets, and comprehensive experiments on benchmark datasets demonstrate that our method achieves clearer segmentation maps without the need for additional annotation, facilitating subsequent polygonization processes. Furthermore, even simple polygonization methods like the Douglas-Peucker algorithm yield satisfactory polygon representations when applied to the segmented outputs generated by our method.</p>