Beyond Gaussians: Fast and High-Fidelity 3D Splatting with Linear Kernels

Haodong Chen,Runnan Chen,Qiang Qu,Zhaoqing Wang,Tongliang Liu,Xiaoming Chen,Yuk Ying Chung

2024-11-19

Abstract: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. The implementation will be made publicly available upon acceptance. \freefootnote{*Corresponding author.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

This paper attempts to solve the problems of blurry artifacts and over - reconstruction in 3D Gaussian Splatting (3DGS), especially the poor performance in high - frequency regions (such as fine textures and details). Specifically, the paper points out:

1. **Limitations of the Gaussian kernel**: The Gaussian kernel has soft - boundary characteristics, which lead to difficulties in separating the foreground and the background, resulting in floating artifacts and over - smoothing. This mixing between neighboring points blurs the detail capture in high - frequency regions and limits the performance of 3DGS in complex scenes.

2. **Deficiencies of existing methods**: Although existing methods have improved the effect of 3DGS by improving the scene structure, enhancing geometric representation, handling blurriness in training images, maintaining rendering consistency, and optimizing density control, there are still obvious artifact problems in high - frequency regions, especially in areas requiring high resolution and sharp transitions.

To solve these problems, the paper proposes **3D Linear Splatting (3DLS)**, which replaces the Gaussian kernel with a linear kernel to achieve clearer and more accurate rendering results, especially in high - frequency regions. Specific contributions include:

- **Introducing 3D Linear Splatting (3DLS)**: Using a linear kernel instead of a Gaussian kernel reduces interference between neighboring primitives, thus achieving sharper transitions and more accurate reconstructions.

- **Proposing Distribution Alignment (DA)**: By adjusting the expansion range of the linear kernel to make it compatible with the Gaussian kernel, seamless integration in the existing framework is ensured, and the reconstruction fidelity is improved.

- **Proposing Adaptive Gradient Scaling (AGS)**: By balancing detail preservation and computational efficiency, the training process is stabilized, enabling 3DLS to effectively capture fine details, sharp transitions, and high - frequency content.

Experimental results show that 3DLS demonstrates state - of - the - art fidelity and accuracy on three benchmark datasets, and achieves a 30% frame rate improvement compared to the baseline 3DGS while maintaining a low memory overhead, making it suitable for real - time applications such as interactive rendering and virtual reality (VR).