CVUSA Research Dataset

Cross-View USA

Image Geolocalization, Autonomous Navigation, Image Synthesis

A large-scale benchmark supporting multiple tasks including image geolocalization, autonomous navigation, and image synthesis, featuring 1.58 million matched ground-to-aerial image pairs.

Matched Pairs

1.58 Million

Locations

879,318

Modalities

3

Citations

640+

Cross-View USA
Download Dataset

Reference Publications

Papers that introduced this dataset

Wide-Area Image Geolocalization with Aerial Reference Imagery

Scott Workman, Richard Souvenir, and Nathan Jacobs

in IEEE International Conference on Computer Vision (ICCV)

Dataset Details

Technical specifications

The Cross-View USA (CVUSA) dataset is a large-scale benchmark for Cross-View Image Geolocalization, Cross-View Image Synthesis, and Autonomous Navigation for UAVs. It provides 1.58 million matched image pairs across the contiguous United States to support research in learning joint semantic representations between ground-level and overhead perspectives. This data enables work in cross-view feature alignment, localization in GPS-denied environments, and generative image synthesis.

Dataset Features

Matched Pairs

  • Scale: 1.58 million matched ground-to-aerial pairs across the US.
  • Locations: 879,318 unique georeferenced locations, including roadway-centered and in-the-wild viewpoints.

Modal Diversity

  • Street View: 1.03 million perspective views extracted from panoramas.
  • Flickr: 551,851 geo-tagged "in-the-wild" social media images.

Multi-Scale Aerial

  • Resolutions: High-fidelity 800x800 tiles at Bing Maps zoom levels 14, 16, and 18.
  • Alignment: Centered at ground query locations for consistent cross-view matching.

Spatial Integrity

  • Sampling: 100x100 grid binned collection strategy to eliminate urban bias.
  • Filtering: CNN-based scene classification to filter for outdoor integrity.