The paper presents a hierarchical 3D Gaussian representation for real-time rendering of very large datasets, addressing the challenge of efficient and high-quality rendering of scenes with thousands of calibrated cameras. The authors introduce a 3D Gaussian Splatting (3DGS) hierarchy that allows for efficient Level-of-Detail (LOD) rendering, enabling real-time navigation in large scenes. The method involves sub subdividing the scene into chunks, optimizing each chunk independently, and then consolidating the hierarchies to form a complete representation. Key contributions include: 1. **Hierarchical 3DGS**: A new hierarchy that allows efficient level selection and interpolation, enhancing visual quality. 2. **Interior Node Optimization**: Methods to optimize the properties of intermediate nodes, improving overall visual quality. 3. **Chunk-Based Training**: A divide-and-conquer approach for training large scenes in independent chunks, allowing parallel processing. The method is demonstrated on several datasets, including one provided by Wayve and three captured using a bicycle helmet-mounted rig. The results show real-time navigation in 3D scenes covering distances up to several kilometers with tens of thousands of images. The paper also discusses related work, implementation details, and evaluation metrics, highlighting the effectiveness of the proposed method in handling large-scale scenes.The paper presents a hierarchical 3D Gaussian representation for real-time rendering of very large datasets, addressing the challenge of efficient and high-quality rendering of scenes with thousands of calibrated cameras. The authors introduce a 3D Gaussian Splatting (3DGS) hierarchy that allows for efficient Level-of-Detail (LOD) rendering, enabling real-time navigation in large scenes. The method involves sub subdividing the scene into chunks, optimizing each chunk independently, and then consolidating the hierarchies to form a complete representation. Key contributions include: 1. **Hierarchical 3DGS**: A new hierarchy that allows efficient level selection and interpolation, enhancing visual quality. 2. **Interior Node Optimization**: Methods to optimize the properties of intermediate nodes, improving overall visual quality. 3. **Chunk-Based Training**: A divide-and-conquer approach for training large scenes in independent chunks, allowing parallel processing. The method is demonstrated on several datasets, including one provided by Wayve and three captured using a bicycle helmet-mounted rig. The results show real-time navigation in 3D scenes covering distances up to several kilometers with tens of thousands of images. The paper also discusses related work, implementation details, and evaluation metrics, highlighting the effectiveness of the proposed method in handling large-scale scenes.