In this paper, we focus on investigating the potential of advanced Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting for 3D scene reconstruction from aerial imagery obtained via sensor platforms with an almost nadir-looking camera. Such a setting for image acquisition is convenient for capturing large-scale urban scenes, yet it poses particular challenges arising from imagery with large overlap, very short baselines, similar viewing direction and almost the same but large distance to the scene, and it therefore differs from the usual object-centric scene capture. We apply a traditional approach for image-based 3D reconstruction (COLMAP), a modern NeRF-based approach (Nerfacto) and a representative for the recently introduced 3D Gaussian Splatting approaches (Splatfacto), where the latter two are provided in the Nerfstudio framework. We analyze results achieved on the recently released UseGeo dataset both quantitatively and qualitatively. The achieved results reveal that the traditional COLMAP approach still outperforms Nerfacto and Splatfacto approaches for various scene characteristics, such as less-textured areas, areas with high vegetation, shadowed areas and areas observed from only very few views.

Whereas emerging learning-based scene representations are predominantly evaluated based on image quality metrics such as PSNR, SSIM or LPIPS, only a few investigations focus on the evaluation of geometric accuracy of the underlying model. In contrast to only demonstrating the geometric deviations of models for the fully optimized scene model, our work aims at investigating the geometric convergence behavior during the optimization. For this purpose, we analyze the geometric convergence of discretized density fields by leveraging respectively derived point cloud representations for different training steps during the optimization of the scene representation and their comparison based on established point cloud metrics, thereby allowing insights regarding which scene parts are already represented well within the scene representation at a certain time during the optimization. By demonstrating that certain regions reach convergence earlier than other regions in the scene, we provide the motivation regarding future developments on locally-guided optimization approaches to shift the computational burden to the adjustment of regions that still need to converge while leaving converged regions unchanged which might help to further reduce training time and improve the achieved quality.