Meta has released a research project which uses AI to find the location of any photographed image. Meta's OrienterNet uses deep learning to determine the accurate position of an image using data from OpenStreetMap. Unlike previous algorithms which have relied on 3D point clouds to identify the locations depicted in images OrienterNet can determine the orientation and position of an image by using OpenStreetMap's 2D maps.
Before reading about how OrienterNet works you probably want to try it out for yourself. You can! Try uploading an image of a street scene to:
OrienterNet Huggingface Demo
I uploaded a Street View screenshot of a random location in San Francisco to the OrienterNet Huggingface demo and it managed to pinpoint the location exactly (the heat map in the screenshot above). At the moment the demo only searches within 512 yards. This means that you have to provide a nearby location for the demo to work.
Presumably the search radius limitations are due to search time constraints. If you download the Jupyter notebook and run the local demo I assume you can change the search radius to any distance you want. However expanding the search radius would increase the time the AI takes to search (and presumably increases the possibility of errors in its location calculations).
At the moment a 512 yard maximum means you are restricted to searching within your local neighborhood. It would be interesting to test OrienterNet's accuracy with a city wide search radius to see how well it can pinpoint images within a whole city.
No comments:
Post a Comment