pcvlab/justzoomin

该数据集伴随论文《Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming》发布，支持跨视图地理定位，即通过地理参考的俯视图像（如卫星或航空影像）来定位地面街景图像。数据集将众包的、有限视野的街景图像与华盛顿特区地区的多尺度卫星/航空瓦片层次结构配对，适用于对基于检索和从粗到细的定位方法进行真实评估。当前版本包含约30万张地面图像，覆盖10公里×10公里的华盛顿特区区域，并附带相应的多尺度俯视瓦片层次结构。数据以TAR分片形式存储，以保持Hugging Face仓库的可管理性。提取后，目录布局与论文中使用的基于路径的PyTorch/tiledwebmaps加载器匹配。This dataset accompanies **Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming**. It supports cross-view geo-localization, where a ground-level street-view image is localized using geo-referenced overhead imagery. The dataset pairs crowd-sourced, limited-field-of-view street-view images with a multi-scale satellite/aerial tile hierarchy over Washington, D.C., making it suitable for realistic evaluation of both retrieval-based and coarse-to-fine localization methods. The current release contains approximately 300k ground-level images over a 10 km × 10 km Washington, D.C. area, together with the corresponding multi-scale overhead tile hierarchy. The data is stored as TAR shards to keep the Hugging Face repository manageable. After extraction, the directory layout matches the path-based PyTorch/tiledwebmaps loader used in the paper.