I'd like to eventually build a system that examines a photograph and estimates where it was taken, similar to OSINT geolocation tools. One approach might be matching the image against a large collection of geotagged Street View or other street-level images. Another might be extracting clues such as a three-way intersection, a nearby factory, and train tracks, then searching map data for locations with those features. Is this realistic for a solo programmer, or does it require the resources of a major company? What subjects and technologies should I learn first?
3 Answers
A solo developer can absolutely build a useful prototype for a town or city. Learn Python, machine learning fundamentals, computer vision, geospatial data, OpenStreetMap querying, embeddings, and vector similarity search. An image-to-text model or vision model can describe the scene, while a retrieval pipeline searches indexed locations. Tools such as RAG and local model runners can help, but they don’t replace the need for a good database. Worldwide coverage at high accuracy is a very different project and would demand huge amounts of imagery, labeling, storage, and compute. Existing OSINT and geospatial search techniques are worth learning too, since many versions of this problem can be solved with structured data and efficient search rather than a single magical AI model.
Start with computer vision, especially convolutional neural networks, image classification, object detection, and transfer learning. Existing open models can identify signs, roads, buildings, vehicles, and other useful clues without training everything from scratch. You could then combine those results with a search system over map or geotagged-image data. A global system trained directly to predict coordinates would require an enormous dataset and substantial GPU time, so a city- or region-sized project is a much more realistic starting point.
The hardest part is probably not the neural network—it’s the data and search problem. You need broad, accurately geotagged imagery from many viewpoints and dates, plus reliable map data. A single photograph is 2D, while the real scene is 3D, and temporary objects, construction, weather, parked vehicles, signs, or billboards can make matching unreliable. For a practical project, break the search into stages: use broad geographic features from map data first, then compare visual details against a smaller set of street images.
Could the system search street-level map imagery within a defined radius after recognizing the main visual features?