A nifty toolkit for geolocation
- Sift: An interface for browsing through details about coordinates and enhancing their metadata with AI(including ML models like GeoCLIP, Visual-Place Recognition, as well as Vision-Language Models like gpt-4o). Design inspired by Elicit.
- Overpass Query: An interface for querying Overpass Turbo with natural language as well as OSM feature and location suggestions. We use RAG + in-context learning to generate better queries, following a similar approach as Overpass NL. Inspired by Bellingcat OSM Search and Cursor.
- Geo Estimation: This runs the SOTA model on geoestimation: GeoClip on user-provided images, and renders a heatmap of predicted results.
- Satellite & Streetview Geolocation: (Experimental as the cold-boot time on Modal is too slow) This matches user target imagery with satellite or streetview imager to determine how similar they are. Uses EigenPlaces for streetview geolocation(ground level) and Sample4Geo for satellite geolocation(crossview).
- Frontend: Next.JS + Vercel
- Database: Redis (will consider postgres for data storage in the future)
- Backend + AI/ML Endpoints: Modal
- Authentication: Clerk
client/: The NextJS codebase of our frontendml/: The codebase of our backend + AI/ML endpoints based on Modal and FastAPI. Uses rye for package management.
Earthkit will be self-hostable. I am still working on documenting the self-hosting process for the time being.