Comparison of 3D splatting with Gaussian and linear kernels. Gaussian kernel-based splatting results in blurred effects, floating artifacts, and over-reconstruction, where small-scale geometry is represented by oversized splats, reducing clarity in high-frequency regions. Panel (a) shows 3D Gaussian Splatting (3DGS), where soft ellipsoid boundaries cause interference between foreground and background. Panel (b) illustrates how the unbounded support of Gaussian kernels hinders separation in 1D distributions. In contrast, panels (c) and (d) show our 3D Linear Splatting (3DLS), where bounded linear kernels reduce interference and enhance separation, achieving clearer and more accurate reconstructions.
Recent advancements in 3D Gaussian Splatting (3DGS) have substantially improved novel view synthesis, enabling high-quality reconstruction and real-time rendering. However, blurring artifacts, such as floating primitives and over-reconstruction, remain challenging. Current methods address these issues by refining scene structure, enhancing geometric representations, addressing blur in training images, improving rendering consistency, and optimizing density control, yet the role of kernel design remains underexplored. We identify the soft boundaries of Gaussian ellipsoids as one of the causes of these artifacts, limiting detail capture in high-frequency regions. To bridge this gap, we introduce 3D Linear Splatting (3DLS), which replaces Gaussian kernels with linear kernels to achieve sharper and more precise results, particularly in high-frequency regions. Through evaluations on three datasets, 3DLS demonstrates state-of-the-art fidelity and accuracy, along with a 30% FPS improvement over baseline 3DGS.
Overview of our method integrated within the 3DGS framework. The process begins with replacing Gaussian kernels with linear kernels to enhance detail capture. Next, DA ensures comprehensive splat coverage, optimizing compatibility with existing frameworks. Finally, AGS is applied to support stable training and improve convergence, resulting in higher visual fidelity.
Evaluation of different kernels on complex patterns to simulate challenging cases. Results indicate that the linear kernel excels in reconstructing high-frequency regions.
Qualitative results demonstrate that our method excels in capturing high-frequency details, fine structures, and sharp transitions, resulting in higher-fidelity reconstructions.
@misc{chen2024linear,
title={Beyond Gaussians: Fast and High-Fidelity 3D Splatting with Linear Kernels},
author={Haodong Chen and Runnan Chen and Qiang Qu and Zhaoqing Wang and Tongliang Liu and Xiaoming Chen and Yuk Ying Chung},
year={2024},
eprint={2411.12440},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2411.12440}
}