SwarmGaussian: Distributed 3D Gaussian Splatting for High Fidelity Aerial Scene Reconstruction

The ability to reconstruct high-fidelity 3D assets from video has long held promise for applications such as inspection, structural assessment, and surveying. While traditional photogrammetry techniques have demonstrated strong performance in controlled settings, they often lack the robustness required to handle complex scene characteristics, including non-Lambertian surfaces, reflective materials, and non-diffuse appearance. Recent neural scene representations have shown significant promise in addressing these challenges. In particular, 3D Gaussian Splatting achieves an effective balance between reconstruction quality and rendering efficiency. However, the explicit and memory-intensive nature of Gaussian primitives makes existing approaches prohibitive for large-scale reconstruction, especially when data is captured across multiple aerial platforms. In this work, we present SwarmGaussian, a framework for incremental and large-scale aerial scene reconstruction from multi-view video captured by swarms of aerial vehicles. The proposed approach integrates semantic scene clustering, depthguided Gaussian optimisation, and spatial partitioning with adaptive levels of detail to enable efficient distributed optimisation and scalable reconstruction. By decomposing complex scenes into semantically and spatially coherent subregions, SwarmGaussian supports parallel reconstruction across multiple agents while maintaining global consistency and high-fidelity visual quality. Experimental results demonstrate that SwarmGaussian enables efficient distributed 3D Gaussian-based reconstruction across aerial swarms, producing high-quality 3D assets suitable for downstream applications such as structural assessment, inspection, and large-scale surveying.