3DCentripetalNet: Building height retrieval from monocular remote sensing imagery

• DOI: 10.1016/j.jag.2023.103311
• Publication Date: 2023-05-16T05:56:18Z
• Authors: Qingyu Li; Lichao Mou; Yuansheng Hua; Yilei Shi; Sining Chen; Yao Sun; Xiao Xiang Zhu
• Keywords: Aerial imagery; Centripetal force; Representation; Monocular

Three-dimensional (3D) building structures are vital to understanding urban dynamics. Monocular remote sensing imagery is a cost-effective data source for large-scale height retrieval when compared LiDAR and multi-view imagery. Existing methods learn footprints maps per pixel via either multi-task network or two separate networks, however, failing consider the information of neighboring pixels that belong identical building. Therefore, we propose learning novel representation 3D buildings, namely centripetal shifts, unified individual instances. Our method termed as 3DCentripetalNet learns shift incorporates planar vertical buildings. Afterward, decoupling module devised corner points. Finally, modeling designed retrieve from learned map We investigate proposed on datasets with different spatial resolutions, i.e., ISPRS Vaihingen dataset (9 cm/pixel) Urban (50 cm/pixel). Experimental results suggest able preserve sharp boundaries, largely alleviate false detections, significantly outperform other competitors. Thus, believe robust solution task monocular