AutoSplat: Constrained Gaussian Splatting
for Autonomous Driving Scene Reconstruction

1University of Toronto, 2Noah's Ark Lab, Huawei Technologies
*Denotes Equal Contribution

AutoSplat can perform realistic simulation of diverse autonomous driving scenarios.

Abstract

Realistic scene reconstruction and view synthesis are essential for advancing autonomous driving systems by simulating safety-critical scenarios. 3D Gaussian Splatting excels in real-time rendering and static scene reconstructions but struggles with modeling driving scenarios due to complex backgrounds, dynamic objects, and sparse views. We propose AutoSplat, a framework employing Gaussian splatting to achieve highly realistic reconstructions of autonomous driving scenes. By imposing geometric constraints on Gaussians representing the road and sky regions, our method enables multi-view consistent simulation of challenging scenarios including lane changes. Leveraging 3D templates, we introduce a reflected Gaussian consistency constraint to supervise both the visible and unseen side of foreground objects. Moreover, to model the dynamic appearance of foreground objects, we estimate residual spherical harmonics for each foreground Gaussian. Extensive experiments on Pandaset and KITTI demonstrate that AutoSplat outperforms state-of-the-art methods in scene reconstruction and novel view synthesis across diverse driving scenarios.

Video

Play with sound.

Motivation

Ground-Truth

AutoSplat motivation.

Unlike 3DGS, AutoSplat reconstructs foreground objects and maintains quality during ego-vehicle lateral shifts.

3DGS - Scene Reconstruction

Ours - Scene Reconstruction

Method

AutoSplat architecture.

AutoSplat reconstructs the background and foreground separately and then fuses them to simulate different scenarios.

Comparison To SOTA

Scene Reconstruction

_________

Novel View Synthesis

Road and Sky Geometry Constraints

AutoSplat enforces road and sky geometry constraints, which maintain view synthesis quality during ego-vehicle lateral shifts.

Reflected Gaussian Consistency

AutoSplat leverages a reflected Gaussian consistency constraint that supervises both the visible and unseen side of foreground objects.

Dynamic Appearance Modelling Via Temporally-Dependent Residual Spherical Harmonics

AutoSplat captures the dynamic appearance of foreground objects by estimating temporally-dependent, residual spherical harmonics.

BibTeX

@misc{khan2024autosplatconstrainedgaussiansplatting,
      title={AutoSplat: Constrained Gaussian Splatting for Autonomous Driving Scene Reconstruction}, 
      author={Mustafa Khan and Hamidreza Fazlali and Dhruv Sharma and Tongtong Cao and Dongfeng Bai and Yuan Ren and Bingbing Liu},
      year={2024},
      eprint={2407.02598},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.02598}, 
}