I have a large set of photographs I made while running. They are geotagged, as I took them with my phone camera. The compass direction is completely unreliable, but lat/lon is more trustworthy. I thought it would be an interesting experiment to extract greenery like grass and trees from these photographs. It can be a useful addition for creating routes that are more pleasant to walk, since the eye-level point of view is not available in OSM. As this is based on my personal photographs, it has the additional benefit of recommending routes that I tend to use. The first challenge I encountered is that out of a few thousand photographs, only a handful were taken during the daytime. After deduplicating and dropping all photos that contain no greenery, this becomes a relatively small set of waypoints. I decided not to extrapolate additional points along OSM ways to keep the dataset small and avoid adding misleading info. The greenery detection works well enough with the SegFormer model, although it is somewhat slow locally. My plan is to select waypoints from this dataset before calling OSRM. This way I get routes that are more enjoyable to walk and run, but are generally longer than the default shortest route. You can find my dataset on Kaggle.