SAMPLE Global 3D Maps | Spatial Models Training Data | 165K Locations | Machine Learning Data | ...

Our dataset delivers unprecedented scale and diversity for geospatial AI training:

🌍 Massive scale: 165,000 unique 3D map sequences and locations, 82,000,000 images, 0.73 PB of Data, orders of magnitude larger than datasets currently used for SOTA Vision/Spatial Models.

⏱️ Constantly growing dataset: 12k new 3D Map sequences and locations monthly.

📷 Full-frame, high-res captures: OVER retains full-resolution, dynamic aspect-ratio images with complete Exif metadata (GPS, timestamp, device orientation), multiple resolutions 1920x1080 - 3840x2880, pre-computed COLMAP poses.

🧭 Global diversity: Environments span urban, suburban, rural, and natural settings across 120+ countries, capturing architectural, infrastructural, and environmental variety.

📐 Rich metadata: Per-image geolocation (±3 m accuracy), timestamps, device pose, COLMAP pose; per-map calibration data (camera intrinsics/extrinsics).

🧠 Applications: Spatial Models Training, Multi-view stereo & NeRF/3DGS training, semantic segmentation, novel view synthesis, 3D object detection, geolocation, urban planning, AR/VR, autonomous navigation.

🤗 1k Scenes Sample: You can access our 1,000-scene sample under the CC-BY-NC license at this link: https://huggingface.co/datasets/OverTheReality/OverMaps_1k

搜集汇总

背景与挑战

背景概述

该数据集提供16.5万个全球3D地图序列及8200万张高分辨率图像，覆盖120多个国家的多样化场景，每月新增1.2万序列。包含完整的地理元数据和多尺度影像，适用于空间建模、3D重建等AI训练任务，并开放1000个场景的CC-BY-NC许可样本。