DARB-Splatting: Generalizing Splatting with Decaying Anisotropic Radial Basis Functions

Following the advent of 3D Gaussian Splatting, this paper generalizes splatting-based 3D reconstruction beyond the exponential family (e.g., Gaussians) by introducing Decaying Anisotropic Radial Basis Functions (DARBFs)—non-negative functions of Mahalanobis distance that retain the practical perks of Gaussian splats for 3D→2D projection. We show DARBFs can approximate the Gaussian’s closed-form integration advantage, expanding the kernel design space. Across several DARBF choices, we observe comparable training convergence and memory footprints to Gaussian splatting, with on-par novel-view quality (PSNR, SSIM, LPIPS), offering a simple, integrable path to more flexible reconstruction kernels.