ES-Gaussian：基于误差空间的高斯溅写映射完成
摘要
准确且经济的室内三维重建对于有效的机器人导航和交互至关重要。传统的 LiDAR 映射提供高精度，但成本高、沉重、耗电，且在新视角渲染方面能力有限。基于视觉的映射虽然成本低且能捕获视觉数据，但常因稀疏点云而难以实现高质量的三维重建。我们提出 ES-Gaussian，一个使用低轨道相机和单线 LiDAR 的端到端系统，实现高质量的室内三维重建。我们的系统特点是视觉误差构建 (VEC)，通过识别和纠正 2D 误差图中几何细节不足的区域来增强稀疏点云。此外，我们引入一种受单线 LiDAR 指导的 novel 3DGS 初始化方法，克服了传统多视角 setup 的局限性，使其在资源受限的环境中能够有效重建。大量实验结果表明，ES-Gaussian 在我们的新 Dreame-SR 数据集和公开数据集上均优于现有方法，尤其在挑战性场景中效果更佳。项目页面地址为 https://chenlu-china.github.io/ES-Gaussian/。
引用
@article{arxiv.2410.06613,
title = {ES-Gaussian: Gaussian Splatting Mapping via Error Space-Based Gaussian Completion},
author = {Lu Chen and Yingfu Zeng and Haoang Li and Zhitao Deng and Jiafu Yan and Zhenjun Zhao},
journal= {arXiv preprint arXiv:2410.06613},
year = {2024}
}
备注
This preprint has been withdrawn due to concerns regarding the originality of certain technical elements, as well as its basis in a company project report that was intended solely for internal discussions. To avoid any potential misunderstandings, we have decided to withdraw this submission from public access. We apologize for any confusion this may have caused