Repolode
Submit
VicenteVivan

VicenteVivan/geo-clip

PyTorch implementation of GeoCLIP, a CLIP-style model that aligns images with GPS coordinates for worldwide image geo-localization.

View on GitHubWebsite
Stars
413
Forks
52
Open issues
9
Last commit
15 days ago

What is it?

GeoCLIP estimates where in the world a photo was taken. It trains an image encoder and a location encoder contrastively on Image-GPS pairs, in the same way OpenAI's CLIP aligns images with text. The model is trained on the MP-16 dataset of about 4.7M geotagged images.

The location encoder treats the Earth as a continuous function, so it can also be used on its own as a pre-trained GPS encoder that produces 512-dimensional embeddings for coordinates. The README reports its use for improving multi-class classification on the Geo-Tagged NUS-Wide dataset when combined with visual features.

The project is installable via pip as the `geoclip` package. Its API provides a `GeoCLIP` class whose `predict` method returns the top-k GPS candidates with probabilities for an image, and a `LocationEncoder` class for embedding coordinates.

Key features

  • 01Predicts top-k GPS coordinates with probabilities for an input image.
  • 02Provides a standalone pre-trained location encoder that embeds latitude/longitude pairs.
  • 03Trained contrastively on the MP-16 dataset of about 4.7M geotagged images.
  • 04Installable via pip or from source.

Use cases

  • Estimating the likely location of photos that lack GPS metadata.
  • Using GPS embeddings as extra features in geo-aware classification models.
  • Researching worldwide image geo-localization or location representation learning.

Quick start

  1. 1

    Install the package from PyPI

    pip install geoclip
  2. 2

    Load the model and predict GPS coordinates for an image

Things to know

  • The README does not state hardware requirements, so GPU needs are unclear.

Topics

#geolocalization#pytorch#contrastive-learning#gps-embeddings#computer-vision#multimodal#deep-learning#geography#geolocation-estimation#machine-learning

FAQ

What is geo-clip?
PyTorch implementation of GeoCLIP, a CLIP-style model that aligns images with GPS coordinates for worldwide image geo-localization. GeoCLIP estimates where in the world a photo was taken. It trains an image encoder and a location encoder contrastively on Image-GPS pairs, in the same way OpenAI's CLIP aligns images with text. The model is trained on the MP-16 dataset of about 4.7M geotagged images.
Is geo-clip free to use? What license does it have?
geo-clip is open source under the MIT license. Check the license terms before using it in commercial products.
Is geo-clip still maintained?
Actively maintained. The last commit to geo-clip was on Oct 7, 2026.
What can I use geo-clip for?
Estimating the likely location of photos that lack GPS metadata. Using GPS embeddings as extra features in geo-aware classification models. Researching worldwide image geo-localization or location representation learning.
How do I install geo-clip?
Install the package from PyPI: pip install geoclip Load the model and predict GPS coordinates for an image
Where can I get geo-clip?
The source code and installation instructions are on GitHub: https://github.com/VicenteVivan/geo-clip

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects