Multimodal Vision Research Laboratory MVRL

Cross-View Time (CVT)

A large dataset containing geotagged ground-level images, corresponding capture time, and co-located overhead images.

Dataset Access

If you want access to the Cross-View Time (CVT) dataset, please submit a request.

Dataset Details

The dataset contains 305,011 ground-level images. Each image is paired with an orthorectified overhead image from Bing Maps (800×800, 0.60 meters/pixel) centered on the capture location, along with the original capture time. It is split into 280,011 training pairs and 25,000 test pairs. Ground images are a subset of the AMOS dataset and phone-camera images from YFCC100M.

This data was collected for research purposes only. Please contact the copyright owners if you need a formal license.

Dataset Citation

If you use this dataset in your research, please cite the following publication(s):

  1. Thumbnail for Learning a Dynamic Map of Visual Appearance
    Salem T, Workman S, Jacobs N. 2020. Learning a Dynamic Map of Visual Appearance. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). DOI: 10.1109/CVPR42600.2020.01245.
    website paper doi bibtex Localization GeoAI

Related Research

  1. Thumbnail for Learning a Dynamic Map of Visual Appearance
    Salem T, Workman S, Jacobs N. 2020. Learning a Dynamic Map of Visual Appearance. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). DOI: 10.1109/CVPR42600.2020.01245.
    website paper doi bibtex Localization GeoAI

Resources

Project Website
More Data: Explore All Datasets →