3D Gaussian Splatting for UAV-Based Reconstruction of Urban Environments

UAV-based 3D reconstruction of large-scale environments has been largely adopted in different domains, starting with virtual reality and 3D documentation and moving on to more practical applications such as land analysis, urban area mapping and disaster management. The reconstruction techniques used are mainly based on two different approaches: with passive sensors, relying on image-only methodologies and with active sensors, which consist of laser-based methods (such as LiDAR). Image-based reconstruction was traditionally based on photogrammetric computer vision, but deep learning has recently innovated these techniques. Deep learning-based methods can learn how to represent three-dimensional scenes to generate realistic renderings from a sequence of images. Among these approaches, Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have recently shown promising results in the generation of accurate rendering, but the quality of the generated 3D point cloud is often neglected. The main object of this research project is to assess the 3D Gaussian Splatting algorithm in terms of quality of the generated 3D reconstruction. In particular, both the original implementation and its updated version, called DN-Splatter, are considered during the evaluation process. DN-splatter takes into consideration normal and depth maps extracted from the data to obtain a more accurate representation from a geometrical point of view. The evaluation is conducted on three different UAV-captured datasets of urban environments, two of them including ground-truth LiDAR point clouds. In order to assess the algorithms' performances, both renders and point clouds are evaluated with appropriate metrics. In particular, PSNR, LPIPS, and SSIM are used to evaluate the generated images, thus the quality of the final renders compared to UAV images. On the other hand, the generated point clouds are evaluated from a geometric point of view, by calculating the point-to-point distance from the LiDAR point cloud. Although the quality of the final renders does not show an evident difference between the two algorithms, the results show DN-Splatter outperforms the original 3D Gaussian Splatting in geometric accuracy. In that regard, the rendering is able to “hide” incorrect 3D reconstructions that are more evident in the point clouds comparison.