VicenteVivan/geo-clip

What it solves

GeoCLIP addresses the difficulty of worldwide image geo-localization—identifying the exact geographical location of a photo based on its visual content. It also provides a way to convert GPS coordinates into semantically rich vector representations (embeddings) that can be used by other AI models.

How it works

Inspired by OpenAI's CLIP, GeoCLIP uses contrastive training on a massive dataset of 4.7 million image-GPS pairs. It aligns visual features from images with geographical locations, treating the Earth as a continuous function to learn distinctive visual patterns associated with specific parts of the globe.

Who it’s for

This tool is designed for researchers and developers building geo-aware neural architectures or applications that need to automatically predict the GPS coordinates of an image.

Highlights

• Worldwide Localization: Achieves state-of-the-art results on multiple benchmark datasets including Im2GPS3k and GWS15k.
• GPS Encoder: Includes a pre-trained location encoder that converts latitude and longitude into 512-dimensional embeddings.
• Downstream Versatility: The learned GPS features can be concatenated with image features to improve multi-class classification accuracy in geo-tagged datasets.