Knowledge distillation-based distributed dynamic 3D Gaussian splatting for large scale scene reconstruction

With the growing demand for applications in fields such as virtual reality, autonomous driving, and urban planning, the efficient and accurate representation and rendering of 3D shapes and scenes have become significant challenges. Recently, 3D Gaussian Splatting (3DGS) has gained widespread attention in academia due to its high-quality novel view rendering and rapid rasterization speed. However, 3DGS still faces several challenges. In large-scale scene reconstruction, centralized methods suffer from poor scalability and the large amount of data imposes a heavy computational burden. In addition, the original 3DGS performs poorly in reconstructing dynamic scenes. To address these issues, we propose a distributed dynamic 3D Gaussian framework based on knowledge distillation. This framework enables multiple clients to collaboratively model 3D a same scene, where local model data obtained by each client is merged and updated on a central server, enhancing scalability and ease of maintenance. Furthermore, since the original 3DGS lacks the spatiotemporal information needed to accurately represent dynamic objects, we introduce time-dependent parameters into the framework to effectively capture the motion of Gaussian points, thereby improving dynamic scene modeling.We evaluate our method on the Waymo and KITTI datasets. Experimental results show that it achieves superior performance in reconstructing urban dynamic scenes.