📍 Try out our demo!

Description

GeoCLIP addresses the challenges of worldwide image geo-localization by introducing a novel CLIP-inspired approach that aligns images with geographical locations, achieving state-of-the-art results on geo-localization and GPS to vector representation on benchmark datasets (Im2GPS3k, YFCC26k, GWS15k, and the Geo-Tagged NUS-Wide Dataset). Our location encoder models the Earth as a continuous function, learning semantically rich, CLIP-aligned features that are suitable for geo-localization. Additionally, our location encoder architecture generalizes, making it suitable for use as a pre-trained GPS encoder to aid geo-aware neural architectures.

Method

Similarly to OpenAI's CLIP, GeoCLIP is trained contrastively by matching Image-GPS pairs. By using the MP-16 dataset, composed of 4.7M Images taken across the globe, GeoCLIP learns distinctive visual features associated with different locations on earth.

📎 Getting Started: API

GeoCLIP requires Python 3.11 or newer and Pillow 13.x. Pillow 13 fixes security vulnerabilities in TIFF decoding and PDF parsing. Python 3.10 users must upgrade Python to use the patched dependency.

You can install GeoCLIP's module using pip:

pip install geoclip

or directly from source:

git clone https://github.com/VicenteVivan/geo-clip
cd geo-clip
python -m pip install .

🗺️📍 Worldwide Image Geolocalization

Usage: GeoCLIP Inference

import torch
from geoclip import GeoCLIP

model = GeoCLIP()

image_path = "image.png"

top_pred_gps, top_pred_prob = model.predict(image_path, top_k=5)

print("Top 5 GPS Predictions")
print("=====================")
for i in range(5):
    lat, lon = top_pred_gps[i]
    print(f"Prediction {i+1}: ({lat:.6f}, {lon:.6f})")
    print(f"Probability: {top_pred_prob[i]:.6f}")
    print("")

🌐 Worldwide GPS Embeddings

In our paper, we show that once trained, our location encoder can assist other geo-aware neural architectures. Specifically, we explore our location encoder's ability to improve multi-class classification accuracy. We achieved state-of-the-art results on the Geo-Tagged NUS-Wide Dataset by concatenating GPS features from our pre-trained location encoder with an image's visual features. Additionally, we found that the GPS features learned by our location encoder, even without extra information, are effective for geo-aware image classification, achieving state-of-the-art performance in the GPS-only multi-class classification task on the same dataset.

Usage: Pre-Trained Location Encoder

import torch
from geoclip import LocationEncoder

gps_encoder = LocationEncoder()

gps_data = torch.Tensor([[40.7128, -74.0060], [34.0522, -118.2437]])  # NYC and LA in lat, lon
gps_embeddings = gps_encoder(gps_data)
print(gps_embeddings.shape) # (2, 512)

Acknowledgments

This project incorporates code from Joshua M. Long's Random Fourier Features Pytorch. For the original source, visit here.

Citation

If you find GeoCLIP beneficial for your research, please consider citing us with the following BibTeX entry:

@inproceedings{geoclip,
  title={GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization},
  author={Vivanco, Vicente and Nayak, Gaurav Kumar and Shah, Mubarak},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}