Summary
- A volunteer team at UC Berkeley have been working on different approaches to automate feature detection and OSM tag extraction from street level imagery.
- Both work tracks built upon a * Panoramax pipeline developed by HeiGIT, which takes an AOI, extracts Panoramax imagery via API, then runs a detection pipeline over the imagery.
- The Computer Division team worked on detection of OpenStreetMap tag
power=polefrom the imagery, with feature annotation sent back to Panoramax. - The Large Language Model team worked on extraction of additional OpenStreetMap tags
surfaceandsmoothnessfor already detected roads, with the goal to push back to OSM.
- The Computer Division team worked on detection of OpenStreetMap tag
- The work will be presented on Thur 21st May via video call (tbc), and subsequent blog. I will update this thread will further details and a link. Feedback and input is really welcomed, to continue this work!
Details
- I work for * HOTOSM as the current Tech Lead. We have collaborated with two student groups at UC Berkley, picking up the work already done by a colleague and volunteer for the osm-tagger project (which uses an LLM to extract road surface and smoothness tags).
- After deliberation and community discussions, we determined:
- Panoramax is the perfect candidate for street level imagery extraction, running through an AI-assisted pipeline for image annotation, then pushing the extracted data back to Panoramax after validation.
- Computer Vision techniques are much more suited to object detection and annotation of Panoramax images. There is some really nice work by Adrien that I’m sure you are all aware of in his detection-tutorial on Gitlab.
- Large Language Models can work quite well to extract additional descriptive tags, such as surface types, smoothness, etc. These tags cannot be stored in Panoramax (as far as I am aware), but instead are suitable for pushing to OpenStreetMap, after feature matching and validation.
The Computer Vision Track: Power Poles
- One of our community members at HOT suggested that detecting power poles could be a very useful and neglected area for OpenStreetMap data, particularly in developing countries.
- The student team at Berkeley spent a lot of time extracting Panoramax images, filtering those with power poles, and tagging the images for training. This dataset will be published and posted here soon, and should hopefully be very valuable for anyone wishing to further refine models.
- The object detection pipeline is generic enough, that in theory we could plug in any object detection model and parameters in future (after training of a model of course).
- There are two possible modes:
- Manual extraction on personal devices. In this case, any one with an area of interest can pull down imagery, run the detection, validate the outputs, then push back to Panoramax. The advantage is this this is targeted for a study area, and uses the local compute available on anyones laptop / device.
- Automated extraction on a server. If the detection is deemed to be accurate enough for a particular feature type, it may be possible to run the pipeline over large areas, annotating features over a certain percentage confidence level.
Large Language Model Track: Road Details
- Through the work done on the osm-tagger project, we determined that detection accuracy of road surface and smoothness is quite accurate, however, the resources required to run such predictions are too high to be viable at scale.
- The primary goal here was the suitable selection of a model that (1) accurately identifies required details (2) is resource efficient enough to make running the detection viable.
- The team settled on the Qwen 2.5B LLM for now, which satisfies the criteria.
- Again, the detection runs as part of a Panoramax-based pipeline for imagery extraction, detection, and output. In this case the output probably does not belong in Panoramax, but instead in OpenStreetMap, once matched up with the feature there.
The end goal is to publish the datasets and detection pipelines as open source repositories, making the workflows easily reproducible, and generally benefiting the quality of Panoramax and OpenStreetMap data 
If anyone is interested in testing this, or picking up the development where it has left off, please get in contact!
More details to come for both projects in the mentioned video call, links, and blog to follow!
Résumé
- Une équipe de bénévoles de l’université de Berkeley travaille actuellement sur différentes approches visant à automatiser la détection des éléments et l’extraction des balises OSM à partir d’images au niveau de la rue.
- Ces deux axes de travail s’appuient sur un pipeline *Panoramax développé par HeiGIT, qui prend une zone d’intérêt (AOI), extrait les images Panoramax via une API, puis exécute un pipeline de détection sur ces images.
- L’équipe de la division Informatique a travaillé sur la détection de la balise OpenStreetMap « power=pole » à partir des images, en renvoyant les annotations des entités à Panoramax.
- L’équipe chargée des grands modèles linguistiques s’est attachée à extraire les balises OpenStreetMap supplémentaires « surface » et « smoothness » pour les routes déjà détectées, dans le but de les renvoyer vers OSM.
- Ce travail sera présenté le jeudi 21 mai lors d’une visioconférence (à confirmer), puis dans un article de blog. Je mettrai à jour ce fil de discussion avec plus de détails et un lien. Vos commentaires et suggestions sont les bienvenus pour nous aider à poursuivre ce travail !