Step-by-Step Guide: How the OpenStreetMap AI Helper Blueprint works

Contributing to OpenStreetMap with the help of AI requires a model trained on an appropriate dataset.

We provide tools and example notebooks to:

• Creating a custom dataset using ground truth data from OpenStreetMap
• Finetuning a YOLO detector using the custom dataset

Once you have a trained model, you can Run Inference in order to find the type of elements you used to train the model (in the example, we chose swimming pools).

Overview

The inference has 4 core stages:

Step 1: Pick a point in the map and download data around it

After a point is selected, a bounding box is computed around it based on the margin argument.

All the existing elements of interest are downloaded from OpenStreetMap using get_elements.

All the tiles are downloaded from MapBox and joined to create a stacked image using download_stacked_image_and_mask. The elements are grouped and converted to a ground truth mask for later usage.

Step 2: Run inference on the stacked image

The stacked image is divided into tiles to run inference using tile_prediction.

For each tile, we run the trained YOLO detector.

If an object of interest is detected, we pass the bounding box to the provided SAM2 model to obtain a segmentation mask.

All the predictions are aggregated into a single stacked output mask.

Step 3: Find existing, new and missed polygons

All the individual mask blobs are converted to polygons for both the stacked output mask and the ground truth mask.

Based on overlap, all the polygons are categorized into existing (green), new (yellow) or missed (red). The really relevant ones are the new (yellow), the others just serve as reference on how the model behaves for polygons already existing in OpenStreetMap.

Step 4: Review, filter and export the new polygons

The new polygons can be manually reviewed and filtered:

The ones you chose to keep will be exportedin OsmChange format.

You can then import the file in any of the supported editors format.

Warning

Make sure to carefully review and edit any predicted polygon.

🎨 Customizing the Blueprint

To better understand how you can tailor this Blueprint to suit your specific needs, please visit the Customization Guide.

🤝 Contributing to the Blueprint

Want to help improve or extend this Blueprint? Check out the Future Features & Contributions Guide to see how you can contribute your ideas, code, or feedback to make this Blueprint even better!