DHGS: Decoupled Hybrid Gaussian Splatting for Driving Scene

Xi Shi,Lingli Chen,Peng Wei,Xi Wu,Tian Jiang,Yonggang Luo,Lecheng Xie

2024-08-18

Abstract:Existing Gaussian splatting methods often fall short in achieving satisfactory novel view synthesis in driving scenes, primarily due to the absence of crafty designs and geometric constraints for the involved elements. This paper introduces a novel neural rendering method termed Decoupled Hybrid Gaussian Splatting (DHGS), targeting at promoting the rendering quality of novel view synthesis for static driving scenes. The novelty of this work lies in the decoupled and hybrid pixel-level blender for road and non-road layers, without the conventional unified differentiable rendering logic for the entire scene. Still, consistency and continuity in superimposition are preserved through the proposed depth-ordered hybrid rendering strategy. Additionally, an implicit road representation comprised of a Signed Distance Function (SDF) is trained to supervise the road surface with subtle geometric attributes. Accompanied by the use of auxiliary transmittance loss and consistency loss, novel images with imperceptible boundary and elevated fidelity are ultimately obtained. Substantial experiments on the Waymo dataset prove that DHGS outperforms the state-of-the-art methods. The project page where more video evidences are given is: <a class="link-external link-https" href="https://ironbrotherstyle.github.io/dhgs_web" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### Problems the paper attempts to solve

This paper aims to solve the problem of insufficient quality in generating new - view images in driving scenarios with existing Gaussian Splatting methods. Specifically, current methods perform poorly in new - view synthesis of static driving scenarios due to the lack of ingenious design and geometric constraints, resulting in less - than - ideal image quality.

#### Main problems
1. **Poor new - view synthesis quality**: Existing Gaussian Splatting methods perform poorly when generating new - view images, especially in crucial perception tasks such as online map drawing.
2. **Discontinuous geometric structure**: Existing methods usually model the entire scene uniformly, which causes the model to be unable to flexibly optimize different types of scene elements (such as roads, buildings, etc.), thus producing discontinuities and artifacts.
3. **Ignoring close - range synthesis quality**: Existing methods often emphasize the quality of overall or distant elements while ignoring the synthesis quality of close - range elements, which is especially important for autonomous driving systems.

### Solutions
To solve the above problems, the authors propose a new neural rendering method - **Decoupled Hybrid Gaussian Splatting (DHGS)**. The main innovations of this method include:

1. **Decoupling scene representation**: The entire driving scene is decoupled into a road layer and a non - road layer, and different Gaussian models are used for modeling respectively. In this way, optimization can be carried out according to the characteristics of each part.
2. **Depth - ordered hybrid rendering strategy**: A depth - sorting - based hybrid rendering strategy is introduced to ensure the consistency and continuity when the road layer and the non - road layer are superimposed.
3. **Implicit road representation**: An implicit road representation based on the Signed Distance Function (SDF) is trained to supervise the geometric properties of the road surface and improve the rendering quality of the road area.
4. **Auxiliary loss functions**: Transmittance loss and consistency loss are introduced to ensure that the generated images have a smooth transition at the boundaries, improving the overall quality and realism of the images.

Through these improvements, DHGS can generate high - quality new - view images in complex driving scenarios, especially performing significantly better than existing methods in the road area. Experimental results show that DHGS outperforms other state - of - the - art methods on the Waymo dataset.

### Summary
This paper solves the quality problems in new - view synthesis of existing Gaussian Splatting methods in driving scenarios by introducing the Decoupled Hybrid Gaussian Splatting method, especially improving the synthesis quality of close - range elements and geometric continuity, providing more realistic and reliable visual input for the simulation and testing of autonomous driving systems.