Title: BEV-GS: Feed-forward Gaussian Splatting in Bird's-Eye-View for Road Reconstruction Show Chinese title

Title: 鸟瞰图中的前馈高斯点云投影：用于道路重建

Abstract: Road surface is the sole contact medium for wheels or robot feet. Reconstructing road surface is crucial for unmanned vehicles and mobile robots. Recent studies on Neural Radiance Fields (NeRF) and Gaussian Splatting (GS) have achieved remarkable results in scene reconstruction. However, they typically rely on multi-view image inputs and require prolonged optimization times. In this paper, we propose BEV-GS, a real-time single-frame road surface reconstruction method based on feed-forward Gaussian splatting. BEV-GS consists of a prediction module and a rendering module. The prediction module introduces separate geometry and texture networks following Bird's-Eye-View paradigm. Geometric and texture parameters are directly estimated from a single frame, avoiding per-scene optimization. In the rendering module, we utilize grid Gaussian for road surface representation and novel view synthesis, which better aligns with road surface characteristics. Our method achieves state-of-the-art performance on the real-world dataset RSRD. The road elevation error reduces to 1.73 cm, and the PSNR of novel view synthesis reaches 28.36 dB. The prediction and rendering FPS is 26, and 2061, respectively, enabling high-accuracy and real-time applications. The code will be available at: \href{https://github.com/cat-wwh/BEV-GS}{\texttt{https://github.com/cat-wwh/BEV-GS}} Show Chinese abstract

Abstract: 路面是车轮或机器人脚部唯一的接触介质。 重建路面对于无人车辆和移动机器人至关重要。 近期关于神经辐射场（NeRF）和高斯点绘法（GS）的研究在场景重建方面取得了显著成果。 然而，它们通常依赖多视角图像输入，并且需要较长的优化时间。 本文提出了一种基于前馈高斯点绘法的实时单帧路面重建方法——BEV-GS。 BEV-GS 包含预测模块和渲染模块。 预测模块遵循鸟瞰图（Bird's-Eye-View）范式，引入了独立的几何网络和纹理网络。 几何参数和纹理参数直接从单帧估计，避免了每场景优化。 在渲染模块中，我们使用网格高斯表示路面并合成新视图，这更符合路面特性。 我们的方法在真实世界数据集 RSRD 上实现了最先进的性能。 路面高度误差降低至 1.73 厘米，新视图合成的 PSNR 达到 28.36 dB。 预测和渲染的 FPS 分别为 26 和 2061，实现了高精度和实时应用。 代码将在以下位置提供： \href{https://github.com/cat-wwh/BEV-GS}{\texttt{https://github.com/cat-wwh/BEV-GS}}