letsGoBlind/Varying_Altitude_Dataset

Varying Altitude Dataset (*Under Review at NeurIPS 2025 data track)

Varying Altitude Dataset is a multi-scale, multi-view image dataset collected using Google Earth Studio. It captures the same geographic locations from three distinct altitude tiers to support research in 3D reconstruction, camera localization, and novel view synthesis. The tiers include:

Satellite View: Overhead ortho-images offering a broad, map-like geodetic context. Aerial View: Captured using a scripted triple-helix camera path that mimics a spiraling drone descent. At each spatial waypoint, three simultaneous images are taken with different camera angles (left, center, right), providing diverse viewpoints and varied baselines. Ground View: Simulated human-level perspectives obtained by virtually walking around the location on accessible paths. Together, these tiers enable cross-altitude and cross-view analysis, making the dataset well-suited for evaluating visual understanding tasks in geographically diverse urban and natural environments.

Each location in our dataset follows a consistent folder structure to support reproducibility and structured parsing. Within each location-specific folder (e.g., IDxxxx, IDxxxx_left, IDxxxx_right, IDxxxx_satellite, or IDxxxx_street), there are three key components:

footage/ — A subfolder containing the full sequence of images captured from Google Earth Studio for that specific viewpoint and trajectory.