Abstract
Recently, 3D Gaussian splatting (3D-GS) has gained popularity in novel-viewscene synthesis. It addresses the challenges of lengthy training times and slowrendering speeds associated with Neural Radiance Fields (NeRFs). Through rapid,differentiable rasterization of 3D Gaussians, 3D-GS achieves real-timerendering and accelerated training. They, however, demand substantial memoryresources for both training and storage, as they require millions of Gaussiansin their point cloud representation for each scene. We present a techniqueutilizing quantized embeddings to significantly reduce memory storagerequirements and a coarse-to-fine training strategy for a faster and morestable optimization of the Gaussian point clouds. Our approach results in scenerepresentations with fewer Gaussians and quantized representations, leading tofaster training times and rendering speeds for real-time rendering of highresolution scenes. We reduce memory by more than an order of magnitude allwhile maintaining the reconstruction quality. We validate the effectiveness ofour approach on a variety of datasets and scenes preserving the visual qualitywhile consuming 10-20x less memory and faster training/inference speed. Projectpage and code is available https://efficientgaussian.github.io