.png){kind=icon}
.png){kind=icon}
Return
Pano2Cloud: enhancing street-view based point cloud by a dual-layer deep matching for 3D building façade reconstruction
.png){kind=icon}
OA
AI
H
K
A
L
G
J
DOI：10.1080/13658816.2026.2682959.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Street-view images (SVI) provide promising data for 3D building façade reconstruction, which commonly relies on multi-view geometric relationships among adjacent images. However, public SVIs are captured with large distances between adjacent viewpoints, which significantly reduces the overlap extent and increases the distortion of the same buildings, leading to unreliable cross-view matching and sparse, incomplete, and unevenly distributed point clouds. To address these challenges, we propose Pano2Cloud, a geospatial inference framework under constrained observation geometry for façade reconstruction. The initial point clouds were estimated by ZoeDepth. To enhance geometric accuracy, the dual-layer deep matching (DLDM) method was designed, in which attention bounding boxes between adjacent SVIs are matched by semantic-spatial clustering in layer I, and feature points are matched by attention-constrained deep-learning network in layer II for point cloud registration. Finally, a contour-driven enrichment process completes the façades by snake algorithm. Pano2Cloud was examined in Wuhan, demonstrating high reconstruction accuracy, with mean cloud-to-cloud distance of 1.85 m, relative geometric errors below 2% and façade reconstruction quality reaching the LoD2 level. Furthermore, the proposed DLDM maintained a high matching rate with increasing baseline distances, while the enrichment process improved building contour completeness to 93.16%.
Keywords:
Street-view panorama
3D building façade reconstruction
dual-layer deep matching
point cloud enhancement
Journal
IF:
5.1
Papers:
2.7K
Citations:
9.3K
Organization
Cited Papers