Scholarlink

In this study, we meticulously analyzed the neural radiation field (NeRF) and three-dimensional Gaussian splatting (3D GS) algorithms, significantly advancing unmanned aerial vehicle (UAV)-based 3D reconstruction technology for large-scale scene modeling. Our research effectively tackles the challenges of high cost, inaccuracy, and inefficiency associated with traditional 3D modeling by introducing implicit representation methods.By presenting a UAV 3D reconstruction dataset with detailed camera parameters and geographic information, we provide a solid foundation for algorithmic evaluation. Our analysis of the Mill 19-Building and “Kaiming-village” datasets reveals that 3D GS outperforms NeRF across key metrics, including the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and learned perceptual image patch similarity (LPIPS). For the Mill 19-Building dataset, 3D GS achieved an SSIM of 0.769, a PSNR of 23.01, and an LPIPS of 0.164, surpassing NeRF's 0.525, 19.54, and 0.512 respectively. Similarly, for “Kaiming-village,” 3D GS scored 0.937 in SSIM, 33.55 in PSNR, and 0.093 in LPIPS, significantly outperforming NeRF's 0.397, 17.87, and 0.766. These metrics underscore 3D GS's superior real-time rendering and depth retention capabilities.These results highlight 3D GS's potential in applications such as urban planning, environmental monitoring, and virtual reality. This research enhances 3D reconstruction technology and expands the horizons for UAV technology applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

3D Reconstruction from Aerial Video Based on Implicit Representation Method

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiaqi Zhou,
  • Yuguang Chen,
  • Anliang Zhu,
  • Wenzhi Tang,
  • Xuefeng Ren,
  • Yong Wang,
  • Shunan Wu,
  • Zhigang Wu

摘要

In this study, we meticulously analyzed the neural radiation field (NeRF) and three-dimensional Gaussian splatting (3D GS) algorithms, significantly advancing unmanned aerial vehicle (UAV)-based 3D reconstruction technology for large-scale scene modeling. Our research effectively tackles the challenges of high cost, inaccuracy, and inefficiency associated with traditional 3D modeling by introducing implicit representation methods.By presenting a UAV 3D reconstruction dataset with detailed camera parameters and geographic information, we provide a solid foundation for algorithmic evaluation. Our analysis of the Mill 19-Building and “Kaiming-village” datasets reveals that 3D GS outperforms NeRF across key metrics, including the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and learned perceptual image patch similarity (LPIPS). For the Mill 19-Building dataset, 3D GS achieved an SSIM of 0.769, a PSNR of 23.01, and an LPIPS of 0.164, surpassing NeRF's 0.525, 19.54, and 0.512 respectively. Similarly, for “Kaiming-village,” 3D GS scored 0.937 in SSIM, 33.55 in PSNR, and 0.093 in LPIPS, significantly outperforming NeRF's 0.397, 17.87, and 0.766. These metrics underscore 3D GS's superior real-time rendering and depth retention capabilities.These results highlight 3D GS's potential in applications such as urban planning, environmental monitoring, and virtual reality. This research enhances 3D reconstruction technology and expands the horizons for UAV technology applications.