Resource-Aware Federated 3D Gaussian Splatting for Large-Scale Scene Reconstruction

Three-dimensional Gaussian splatting (3D-GS) has emerged as a promising technique for large-scale scene reconstruction due to its high rendering efficiency and fidelity. However, training large-scale 3D-GS models at wireless networks faces significant challenges due to limited communication, computation, and graphics processing unit (GPU) memory resources. This paper investigates a novel resource-aware federated learning framework for large-scale 3D-GS model training under such constraints. Specifically, we develop a latency- and memoryaware model lightweighting mechanism that adaptively selects and prunes Gaussian points to balance rendering quality and training overhead. Specifically, we quantitatively evaluate the importance of different Gaussian points and exploit a novel importance-to-latency ratio criterion to determine the optimal number of pruned points under memory and communication latency constraints. Extensive experiments on large-scale scenes show that the proposed design significantly accelerates convergence, maintains high rendering quality, and reduces training latency compared to state-of-the-art federated 3D-GS baselines.