I'd like to eventually build a system that examines a photo and estimates where it was taken, using an OSINT-style approach. One possible method would be comparing the image against a large collection of geotagged Street View or other street-level photos. Another idea is to extract clues—such as a three-way intersection, a nearby factory, or train tracks—and search map data such as OpenStreetMap for matching locations. Is this something a solo programmer could realistically learn and build, perhaps starting with one town or city, or does it require the kind of data and computing resources only very large companies have?
4 Answers
The computer-vision topic to study is convolutional neural networks, along with object detection, image embeddings, and transfer learning. You probably would not train a giant model from scratch; an existing vision model could identify signs, roads, buildings, vehicles, and other useful clues. The difficult part is getting a large, accurate collection of geotagged images and enough GPU time to train and evaluate the system. A city-scale prototype is realistic for one person, but worldwide coverage is a much larger engineering and data problem.
A practical design would split the task into stages. First, use a vision model to extract visual clues and perhaps create an embedding for the image. Next, search a geospatial database for broad features—road layouts, waterways, railways, building footprints, factories, and intersections—before comparing against street-level imagery. An image is only a 2D view of a changing 3D scene, so occlusion, construction, weather, camera angle, moving vehicles, and misleading signs can all produce false matches.
This is less like a magic chatbot and more like a large-scale search and ranking system. You need clean geospatial data, efficient indexing, image matching, and a way to compare likely candidates. Human-labeled training data may involve millions of examples, and licensing or access to commercial street imagery can be a major limitation. Existing OSINT and visual-search tools may already solve parts of this, so learning those techniques can be useful before building a custom model.
A solo developer can absolutely build a useful proof of concept for a neighborhood, town, or possibly a city. Start with Python, basic machine learning, computer vision, geospatial databases, OpenStreetMap data, and pretrained models. You can later explore image retrieval, vector databases, and retrieval-augmented systems that let a language model organize search results. A reliable global system would require enormous coverage, data maintenance, compute, and testing—resources closer to those of a major company or research project.
That makes sense. Starting with broad map features instead of searching every street-level image should make the problem much more manageable.