Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GGS: Generalizable Gaussian Splatting for Lane Switching in Autonomous Driving

Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, Chunxia Xiao

TL;DR

This work tackles the challenge of realistic, lane-switching novel-view synthesis for autonomous driving when multi-lane data are unavailable. It introduces GGS, a framework that combines multi-view depth refinement, virtual lane generation, and diffusion-based supervision within a generalized 3D Gaussian Splatting pipeline to render views from unseen lanes. Key contributions include a virtual lane module, a multi-lane diffusion loss, and depth refinement that together achieve state-of-the-art results on KITTI and BrnoUrban without LiDAR. The approach enables effective simulation-based testing for autonomous driving systems and improves generalization to new road scenarios, with practical impact for robust traffic scenario rendering and lane-change testing.

Abstract

We propose GGS, a Generalizable Gaussian Splatting method for Autonomous Driving which can achieve realistic rendering under large viewpoint changes. Previous generalizable 3D gaussian splatting methods are limited to rendering novel views that are very close to the original pair of images, which cannot handle large differences in viewpoint. Especially in autonomous driving scenarios, images are typically collected from a single lane. The limited training perspective makes rendering images of a different lane very challenging. To further improve the rendering capability of GGS under large viewpoint changes, we introduces a novel virtual lane generation module into GSS method to enables high-quality lane switching even without a multi-lane dataset. Besides, we design a diffusion loss to supervise the generation of virtual lane image to further address the problem of lack of data in the virtual lanes. Finally, we also propose a depth refinement module to optimize depth estimation in the GSS model. Extensive validation of our method, compared to existing approaches, demonstrates state-of-the-art performance.

GGS: Generalizable Gaussian Splatting for Lane Switching in Autonomous Driving

TL;DR

This work tackles the challenge of realistic, lane-switching novel-view synthesis for autonomous driving when multi-lane data are unavailable. It introduces GGS, a framework that combines multi-view depth refinement, virtual lane generation, and diffusion-based supervision within a generalized 3D Gaussian Splatting pipeline to render views from unseen lanes. Key contributions include a virtual lane module, a multi-lane diffusion loss, and depth refinement that together achieve state-of-the-art results on KITTI and BrnoUrban without LiDAR. The approach enables effective simulation-based testing for autonomous driving systems and improves generalization to new road scenarios, with practical impact for robust traffic scenario rendering and lane-change testing.

Abstract

We propose GGS, a Generalizable Gaussian Splatting method for Autonomous Driving which can achieve realistic rendering under large viewpoint changes. Previous generalizable 3D gaussian splatting methods are limited to rendering novel views that are very close to the original pair of images, which cannot handle large differences in viewpoint. Especially in autonomous driving scenarios, images are typically collected from a single lane. The limited training perspective makes rendering images of a different lane very challenging. To further improve the rendering capability of GGS under large viewpoint changes, we introduces a novel virtual lane generation module into GSS method to enables high-quality lane switching even without a multi-lane dataset. Besides, we design a diffusion loss to supervise the generation of virtual lane image to further address the problem of lack of data in the virtual lanes. Finally, we also propose a depth refinement module to optimize depth estimation in the GSS model. Extensive validation of our method, compared to existing approaches, demonstrates state-of-the-art performance.
Paper Structure (16 sections, 16 equations, 8 figures, 3 tables)

This paper contains 16 sections, 16 equations, 8 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. 3D Gaussian Splatting
  4. Generalized model
  5. Diffusion Model
  6. Methodology
  7. Background
  8. Mutli-View Depth Refinement Module
  9. Virtual Lane Generation Module
  10. Multi-Lane Diffusion Loss
  11. Loss Function
  12. Experiments
  13. Evaluation on KITTI and BrnoUrban
  14. Assessing Cross-dataset Generalization
  15. Ablation Study
  16. ...and 1 more sections

Figures (8)

  • Figure 1: Our GGS method can achieve high-quality lane switching in autonomous driving scenarios.
  • Figure 2: The overall framework of the GGS. Input multiple frames and estimate depth maps through MVS and multi view depth refinement modules, combined with 3DGS to synthesize novel views. And through the virtual lane generation module, switch lanes with high quality. In addition, multi-lane diffusion loss is introduced to supervise the novel view synthesis in the presence of obstacles.
  • Figure 3: The method of using a lane converter to create a virtual lane and then switching back to the real lane enables the model to improve the quality of lane switching.
  • Figure 4: If we switch from the right lane to the left lane, the red area represents the blind spot of the right lane. When rendering the left lane, in order to avoid voids, the content of that area needs to be imagined in some way.
  • Figure 5: Comparison results of novel view synthesis based on KITTI for residential, road, and urban scenes.
  • ...and 3 more figures