2026-10-11T16:26:42Z https://www.scidb.cn/oai

10.57760/sciencedb.012jp 2026-09-20T10:18:18Z

2026-09-20 VSG-GS: UAV Oblique Imagery for Semantic-Guided 3D Gaussian Splatting Bridge Reconstruction doi:10.57760/sciencedb.012jp en This dataset accompanies the VSG-GS framework, comprising three components: (1) raw UAV oblique imagery, (2) source code, and (3) workflow documentation.1. Data Generation and AcquisitionThe raw images were captured using a DJI Mavic 3 Air quadcopter UAV equipped with a 20 MP CMOS camera (focal length: 24 mm equivalent). Two bridge sites were surveyed:Tian'e Longtan Bridge (N 25°02′–25°04′, E 107°08′–107°10′), located in Tian'e County, Hechi City, Guangxi Zhuang Autonomous Region, China. This is a long-span concrete arch bridge under construction spanning a river valley surrounded by mountainous terrain. Data were collected in December 2024. The UAV flew at an altitude of 80–120 m above the launch point, with both nadir and oblique (45°) camera angles, yielding 1,247 images at 5,280 × 3,956 px resolution. Ground sampling distance (GSD) is approximately 2.0–3.5 cm/pixel.Caiyuanba Bridge (N 29°31′–29°33′, E 106°32′–106°34′), located in Yuzhong District, Chongqing, China. This is a steel-tied arch bridge crossing the Yangtze River in a dense urban setting. Data were collected in March 2025. The flight was conducted at 100–150 m altitude with multi-angle oblique photography, producing 892 images at 5,280 × 3,956 px resolution. GSD is approximately 2.5–4.0 cm/pixel.All images were captured in JPEG format. Invalid images (severe motion blur, exposure anomalies, or excessive redundancy) were manually removed during post-screening.2. Tools and SoftwareCOLMAP (v3.9.1): Structure-from-Motion for camera pose recovery and sparse point cloud initialization.Florence-2 (Microsoft, pretrained): Zero-shot box-level bridge localization using the text prompt "The most prominent bridge in the scene."SAM2 (Meta, fine-tuned as Bridge-SAM2): Promptable segmentation for pixel-level bridge mask generation. The fine-tuning was performed on a manually annotated subset (~200 images per dataset) with iterative hard-sample refinement over 3 rounds.3D Gaussian Splatting (base implementation by Kerbl et al., 2023): Modified with mask-constrained loss and adaptive density control restricted to bridge ROI.Hardware: Training was conducted on an NVIDIA RTX 4090 GPU (24 GB VRAM) with an AMD Epyc 7542 CPU and 512 GB RAM. 3D Gaussian Splatting; UAV oblique imagery; bridge reconstruction; semantic-guided reconstruction; vision-language model; digital twin 左恒滔 EMBARGO https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank