SplatAnything applies Gaussian Spacing Normalization to enhance feed-forward 3D Gaussian Splatting across diverse scenes and viewpoints.
A new AI research paper titled "SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes" has been published on arXiv, introducing a novel approach to improve 3D Gaussian Splatting (3DGS). The research, led by Hao Yu, Zhenwei Wang, Yongzhen Hu, Yuning Gong, and Junta Wu, addresses key limitations in existing 3DGS methods by proposing Gaussian Spacing Normalization (GSNorm).
Traditional 3DGS methods rely on radial-distance-normalized spaces, which often result in varying spacing between Gaussian centers across different scene scales and camera configurations. This inconsistency leads to ambiguous supervision and systematic shifts in Gaussian-scale targets. The SplatAnything framework introduces GSNorm to canonicalize scene-level Gaussian spacing, providing a consistent spatial reference for Gaussian predictions. This innovation enhances the robustness of 3D reconstruction under varying scene scales and viewpoints.
The researchers established a systematic benchmark to evaluate SplatAnything's performance across diverse datasets, input sparsity levels, and challenging extrapolated viewpoints. Experimental results demonstrate that SplatAnything achieves state-of-the-art performance, with significant improvements under substantial variations in scene scale and camera configurations.
SplatAnything, the third iteration of the InfiniSplat series, is designed for applications such as autonomous driving, robot navigation, augmented reality (AR), virtual reality (VR), and game development. It supports advanced features like Rig-Recon for multi-camera setups and Pano-Recon for 360° panoramas. The tool is open-source and available on GitHub at https://github.com/zju3dv/SplatAnything.
4+ with CUDA compatibility. Users can install the tool by cloning the repository and installing the required dependencies.
Source link
