3D Skew Gaussian Splatting

3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction and rendering, enabling photorealistic novel view synthesis from sparse images. However, the fundamental reliance on 3D Gaussian distributions struggles to capture shape and color discontinuities in real-world scenes, such as sharp edges and intricate surface details. In this paper, we introduce 3DSGS (3D Skew Gaussian Splatting), a novel framework that significantly enhances the accuracy and efficiency of Gaussian- based neural scene representation. Our key insight lies in extending the original 3D Gaussian to its more general Skew Gaussian counterpart, where the original Gaussian distribution emerges as a special symmetric case. This generalized primitive inherits all desirable properties of Gaussians while gaining intrinsic asymmetric modeling capability, enabling more accu- rate representations for complex scenes. Then we propose a refined representation for the opacity parameter, increasing the flexibility of each skew Gaussian to better handle transparency and fine structures. Furthermore, we introduce a depth-aware densification criterion that intelligently manages the creation, splitting, and cloning of skew Gaussians. Extensive experiments demonstrate that 3DSGS achieves superior rendering quality, particularly in regions with intricate details, transparency, and view-dependent effects with real-time rendering speed. The code will be open-sourced upon paper acceptance.