Skip to content

Implementation of 'Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization' [CVPR 2021]

License

Notifications You must be signed in to change notification settings

aysim/comingdowntoearth

Repository files navigation

Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization

Implementation of the CVPR 2021 paper Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization.

Abstract

The goal of cross-view image based geo-localization is to determine the location of a given street view image by matching it against a collection of geo-tagged satellite im- ages. This task is notoriously challenging due to the drastic viewpoint and appearance differences between the two do- mains. We show that we can address this discrepancy ex- plicitly by learning to synthesize realistic street views from satellite inputs. Following this observation, we propose a novel multi-task architecture in which image synthesis and retrieval are considered jointly. The rationale behind this is that we can bias our network to learn latent feature rep- resentations that are useful for retrieval if we utilize them to generate images across the two input domains. To the best of our knowledge, ours is the first approach that cre- ates realistic street views from satellite images and local- izes the corresponding query street-view simultaneously in an end-to-end manner. In our experiments, we obtain state- of-the-art performance on the CVUSA and CVACT bench- marks. Finally, we show compelling qualitative results for satellite-to-street view synthesis.

The code has been implemented & tested with Python 3.6.9 and Pytorch 1.5.0.

Usage

Datasets and Polar transformation

Pretrained Models

Train

  • For CVUSA, use train_synthesis_cvusa.py
  • For CVACT, use train_synthesis_cvact.py

Test

To test our architecture use the pretained models, given above, and run cvusa_test.py and cvact_test.py.

Cite

If you use our implementation, please cite:

@InProceedings{Toker_2021_CVPR,
    author    = {Toker, Aysim and Zhou, Qunjie and Maximov, Maxim and Leal-Taixe, Laura},
    title     = {Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {6488-6497}
}

About

Implementation of 'Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization' [CVPR 2021]

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages