Multi-Res-3DGS: Multi-Resolution 3d Gaussian Splatting Bound with a Subdivided Mesh Sequence

Multi-resolution functionality is a powerful tool in video content delivery services, ensuring compatibility across various devices and optimizing bandwidth usage. However, 3D Gaussian Splatting (3DGS), an explicit radiance field for new view synthesis and efficient rendering, currently lacks such functionality. This paper introduces multi-resolution functionality for 3DGS for the first time. Recognizing that the number of 3D Gaussians is a critical factor influencing resource consumption in 3DGS rendering, we propose a method to control this number across various resolutions. We propose a novel pyramidal data structure called Multi-Res-3DGS, which binds 3D Gaussians with a sequence of subdivided meshes. In this framework, the pyramidal meshes guide the multi-resolution through iterative subdivision, while the 3D Gaussians encode details associated with the mesh faces. The Multi-Res-3DGS can be trained using a combination of existing techniques and rendered at the best resolution according to the available resources. Evaluations using the NeRF-Synthetic dataset demonstrate that our approach realizes the multi-resolution functionality with minimal PSNR or LPIPS losses.