OpenStreetView-5M
收藏资源简介:
OpenStreetView-5M (OSV-5M) 是由 LIGM/ENPC/UGE 实验室创建的全球规模、开放获取的街景图像数据集,旨在推动现代计算机视觉方法在全局视觉地理定位领域的研究与应用。该数据集包含超过510万张地理参考的街景图像,覆盖225个国家和地区,训练集包含4894685张图像,测试集包含210122张图像,图像高度为512像素,平均宽度为792±127像素。数据来源于 Mapillary 平台,该平台汇集了全球用户的众包街景图像,并遵循 CC BY-SA 许可证,允许免费使用并标注来源。 在创建过程中,研究者们首先在 Mapillary 平台的18亿张图像中,通过定义100×100米的全球网格,随机选择每个单元格中的一张图像,然后根据局部图像密度的负0.75次方权重进行训练集和测试集的采样。此外,还通过手工和学习型过滤器去除低质量图像,如模糊、曝光不当、紫色调异常等,并确保训练集和测试集在空间上至少相隔1公里,避免过度依赖于特定地点的特征。 OSV-5M 数据集广泛应用于计算机视觉领域的多个任务,如全球视觉地理定位、自监督学习和生成建模等。它为研究者提供了一个标准化的基准,用于评估和改进模型在地理特征学习和位置预测方面的性能,推动了在新闻、取证、历史与文化研究等领域的应用发展。数据集可通过 huggingface.co 平台访问,为全球研究者提供了一个开放且易于获取的资源.
The OpenStreetView-5M (OSV-5M) is a globally scaled, open-access street view image dataset created by the LIGM/ENPC/UGE laboratory, aimed at advancing research and application of modern computer vision methods in the field of global visual geolocation. The dataset comprises over 5.1 million georeferenced street view images, covering 225 countries and regions, with the training set containing 4,894,685 images and the test set containing 210,122 images. The images have a height of 512 pixels and an average width of 792±127 pixels. The data originates from the Mapillary platform, which aggregates crowd-sourced street view images from global users and is licensed under CC BY-SA, allowing free use with attribution. During the creation process, researchers first randomly selected one image from each cell of a 100×100-meter global grid from the 1.8 billion images on the Mapillary platform. The training and test sets were then sampled based on the negative 0.75 power of local image density. Additionally, low-quality images, such as blurred, improperly exposed, or abnormally purple-toned images, were manually and algorithmically filtered out. It was also ensured that the training and test sets were spatially separated by at least 1 kilometer to avoid over-reliance on specific location features. The OSV-5M dataset is widely used in various tasks within the field of computer vision, such as global visual geolocation, self-supervised learning, and generative modeling. It provides researchers with a standardized benchmark for evaluating and improving model performance in geographical feature learning and location prediction, promoting the development of applications in fields such as journalism, forensics, historical, and cultural studies. The dataset is accessible through the huggingface.co platform, offering a freely and easily accessible resource to researchers worldwide.
OpenStreetView-5M 数据集概述
基本信息
- 数据集名称: OpenStreetView-5M
- 简介: 首个大规模开放的街景图像地理定位基准数据集。
- 项目页面: https://imagine.enpc.fr/~ioannis.siglidis/osv5m/
- 论文地址: https://arxiv.org/abs/2404.18873
- 演示地址: https://huggingface.co/spaces/osv5m/plonk
数据集详情
- 托管地址: https://huggingface.co/datasets/osv5m/osv5m
- 下载方式: 运行
python scripts/download-dataset.py - 导入方式: 参见 DATASET.md
模型信息
-
模型使用示例: python from PIL import Image from models.huggingface import Geolocalizer
geolocalizer = Geolocalizer.from_pretrained(osv5m/baseline) img = Image.open(.media/examples/img1.jpeg) x = geolocalizer.transform(img).unsqueeze(0) gps = geolocalizer(x)
-
模型评估: 运行
python evaluation.py exp=eval_best_model dataset.global_batch_size=1024
实验复现
-
环境安装: bash conda env create -f environment.yaml conda activate osv5m
-
预处理: 运行
python scripts/preprocessing/preprocess.py data_dir=datasets do_split=1000 -
训练最佳模型: 运行
python train.py exp=best_model computer.devices=1 computer.num_workers=16 dataset.global_batch_size=2
引用
bibtex @article{osv5m, title = {{OpenStreetView-5M}: {T}he Many Roads to Global Visual Geolocation}, author = {Astruc, Guillaume and Dufour, Nicolas and Siglidis, Ioannis and Aronssohn, Constantin and Bouia, Nacim and Fu, Stephanie and Loiseau, Romain and Nguyen, Van Nguyen and Raude, Charles and Vincent, Elliot and Xu, Lintao and Zhou, Hongyu and Landrieu, Loic}, journal = {CVPR}, year = {2024}, }

- 1OpenStreetView-5M: The Many Roads to Global Visual Geolocation · 2024年



