ZeroVO
收藏资源简介:
ZeroVO是一个创新的视觉里程计(VO)算法,能够在多种相机和环境之间实现零样本泛化,克服了现有方法对预定义或静态相机校准设置的依赖。该算法包含三个主要创新:首先,设计了一种无需校准的、几何感知的网络结构,能够处理估计深度和相机参数中的噪声;其次,引入了一种基于语言的先验知识,通过注入语义信息来增强特征提取和泛化到之前未见过的领域;第三,开发了一种灵活的半监督训练范式,利用未标记数据迭代适应新场景,进一步提高了模型在多种现实世界场景中的泛化能力。该研究分析了复杂的自动驾驶环境,在三个标准基准测试(KITTI、nuScenes和Argoverse 2)以及一个新引入的高保真合成数据集(来自Grand Theft Auto)上,比先前的方法提高了30%以上。由于无需微调或相机校准,我们的工作扩大了VO的应用范围,为大规模的实际部署提供了一个通用的解决方案。
ZeroVO is an innovative visual odometry (VO) algorithm that enables zero-shot generalization across diverse cameras and environments, eliminating the reliance of existing methods on predefined or static camera calibration settings. The algorithm encompasses three core innovations: First, a calibration-free geometry-aware network architecture is designed to handle noise in both estimated depth and camera parameters. Second, a language-based prior knowledge is introduced to enhance feature extraction and generalization to previously unseen domains by injecting semantic information. Third, a flexible semi-supervised training paradigm is developed, which leverages unlabeled data to iteratively adapt to new scenes and further improves the model’s generalization capability across various real-world scenarios. This study analyzes complex autonomous driving environments, and achieves over 30% performance improvement over prior state-of-the-art methods on three standard benchmarks (KITTI, nuScenes, and Argoverse 2) as well as a newly introduced high-fidelity synthetic dataset from Grand Theft Auto. Since no fine-tuning or camera calibration is required, our work expands the application scope of VO and provides a universal solution for large-scale real-world deployment.
ZeroVO: Visual Odometry with Minimal Assumptions 数据集概述
基本信息
- 作者: Lei Lai, Zekai Yin, Eshed Ohn-Bar (波士顿大学)
- 发表会议: CVPR 2025
- 项目页面: https://zvocvpr.github.io/
- 数据集类型: 视觉里程计(Visual Odometry)
- 核心创新: 零样本泛化能力,无需预定义相机标定
关键贡献
-
标定无关的几何感知网络
- 无需预定义标定
- 可处理深度估计和相机参数噪声
-
基于语言的语义先验
- 通过语言模型注入语义信息
- 增强特征提取和跨域泛化能力
-
灵活的半监督训练
- 利用未标注数据迭代适应新场景
- 提升现实场景泛化能力
GTA V数据集特性
- 数据规模: 922段驾驶序列(每段25秒,10FPS)
- 场景多样性:
- 多种天气条件(雪、雨、昼夜)
- 不同驾驶速度(含高速机动)
- 复杂交通场景和动态物体
- 相机设置变化和镜头畸变效果
- 越野场景(山地、沙漠、森林)
实验性能
定量结果(误差指标)
| 方法 | KITTI (tₑᵣᵣ) | nuScenes (ATE) | Argoverse (sₑᵣᵣ) | GTA (rₑᵣᵣ) |
|---|---|---|---|---|
| XVO | 16.82 | 8.30 | 0.12 | 12.64 |
| M+DS | 14.22 | 10.46 | 0.13 | 10.38 |
| ZeroVO | 7.69 | 6.79 | 0.11 | 10.63 |
| ZeroVO+ | 6.81 | 6.03 | 0.09 | 7.99 |
| LiteZeroVO+ | 8.85 | 6.87 | 0.11 | 12.16 |
模型变体
- ZeroVO+: 半监督训练+多模态伪标签选择(0.6 FPS)
- LiteZeroVO+: 资源优化版本(5 FPS)
资源链接
- 论文: [PDF]
- 代码: [GitHub]
- 数据集: [GTA Dataset]
- 补充视频: [Supplementary Video]
引用格式
bibtex @inproceedings{donotciteplaceholder, title={{ZeroVO}: Visual Odometry with Minimal Assumptions}, author={Lai, Lei and Yin, Zekai and Ohn-Bar, Eshed}, booktitle={CVPR}, year={2025} }




