Map Features in OpenStreetMap with Computer Vision
blog.mozilla.ai
blog.mozilla.ai
The algorithms have problems with false positives, and with mapping straight or rectangular objects as wobbly, as shown in the second-to-last screenshot.
As a helper to detect missing features, this is a precious tool. But we still need human intervention to make sure the detected objects are drawn correctly.
See also: https://wiki.openstreetmap.org/wiki/Import/Guidelines and https://wiki.openstreetmap.org/wiki/Automated_Edits_code_of_...
Hi, I am the author.
The demo app and any provided code example includes a step asking a human to verify the detected features. You can't upload them automatically unless you modify the source code.
I reiterate the human verification across the docs, linked post, and any code samples.
I haven't ever uploaded features automatically. In fact I manually edited and labeled hundreds of swimming pool samples myself before even training the first version.
Happy to hear and implement any ideas on how to improve the process to prevent automated features to be uploaded.
I know some people might say: just don't publish the tool, I think we can do better at embracing AI and having an open discussion.
Your docs show a simple image where the user can choose to keep a new object or not. [0] Afterwards it says: "The ones you chose to keep will be uploaded to OpenStreetMap using upload_osm.". This is uploading features automatically. The fact that it asks 'are you sure' is just silly. We all know if humans have to click yes 90% of the time, and no 10% of the time, they'll miss a lot of no's.
The image also proofs that:
- You don't see any polygons properly. You just see a an image of where the pool is. Already on the image I can see that if the polygons align to that image, it will be a total mess.
- You don't see any polygons further away from the object.
Both these points are in stereo's reply that the resulted data was a mess.
Please consider pulling the project. This will generate a lot of data that volunteers will have to check and revert.
[0] https://github.com/mozilla-ai/osm-ai-helper/blob/main/docs/s...
This is just not true. The data can be easily identified with the `created_by` tag. And I have been reviewing myself any data uploaded with the demo (with a clear different criteria on what is good enough)
We all strive for better open data. I upstream feel there is a risk that automated uploads could be easier with this project, creating more boring work for them which is already enough of a problem, that animosity will be a net negative for everyone in this space. Technical solutions such as new tags or opt out schemes will not solve the problem.
Idea: do not automatically create features that a human can simply approve, instead require them to draw the polygon themselves.
You acknowledge the problem of AI slop up-front, but seem to have chosen to plow forward anyway. Please do consider your actions and their effects carefully [1]. You are in a position of potential influence; try not to squander it.
All the best,
-HG
You can disagree on whether the measures are effective, of course, but they're clearly not thoughtless.
> The polygons the algorithm had drawn were consistently of poor quality with stray nodes and nodes far outside the pool boundaries, and the imports hadn't been discussed with local communities.
Otherwise, you are discounting their effort based on a prejudice — others might be unable to supervise an AI, but someone who's actually developed it might have a better chance of success.
This is a because the polygon is drawn as a mask in order to overlay it on the image. The actual polygon being uploaded doesn't have the wobbly features.
It is True there are cases were the predicted polygon is wobbly and I encourage people to discard them. However I didn't publish this demo until I got a first version of the model that reached some minimum quality.
There is logic in the code to simplify the shape of the predicted polygon in order to avoid having too many nodes.
I have disabled the hosted demo for now, and will remove the uploading part from the code in favor of showing an URL that will open the editor at the location.
If its of any help, you can find any contributed polygon with the tag `created_by=https://github.com/mozilla-ai/osm-ai-helper`. Feel free to remove all of them (or I can do it myself once I access a PC).
I will be happy to continue the discussion on what is a good prediction or not. I have mapped a lot of swimming pools myself and edited and removed a lot of (presumably) human contributed polygons that looked worse (too my eyes) than the predictions I approved to be uploaded.
Hi, I didn't know about this possibility. I should have better researched what were the different options. I will be taking a look on implementing this approach.
Something else you need to be mindful of is that the mapbox imagery may be out of date, especially for the super zoomed in stuff (which comes from aerial flights). So e.g., a pool built 2 years ago might not show up.
I've spent a lot of time building models for tree mapping. In theory you could use that as a pipeline with OAM to generate forest regions for OSM and it would probably be better than human labels which tend to be very coarse. I wouldn't discount AI labeling entirely, but it does need oversight and you probably want a high confidence threshold. One other thought is you could compare overlap between predicted polygons and human polygons and use that as a prompt to review for refinement. This would be helpful for things like individual buildings which tend to not be mapped particularly well (i.e. tight to the structure), but a modern segmentation model can probably provide very tight polygons.
It's useful for finding ones that haven't been mapped but not for drawing them. It can get the 4 corners pretty accurate for pools that are square, many are half round at the ends though
As disclaimed in the demo and code, the example model was trained only with data from Galicia on a Google Colab. A robust enough models would require more data and compute.
> it's definitely uploading crap.
What was uploaded was what a human approved.
> It's useful for finding ones that haven't been mapped but not for drawing them. It can get the 4 corners pretty accurate for pools that are square, many are half round at the ends though
I couldn't dedicate enough time on the best way to refine the predictions, but happy to hear and discuss any ideas.
Ideas I have are:
- Try an oriented bounding box model instead of detection + segmentation. It will not be useful for not square shapes but will definitely generate more accurate predictions. - Build some sort of https://es.wikipedia.org/wiki/RANSAC that tries to fits rectangles and/or other shapes as an step to postprocess the predicted mask.
Yes, I hit approve on the best one because I was curious to see the actual final polygon. (I then went and fixed it.) You wrote above / I was responding to:
>> This is a because the polygon is drawn as a mask in order to overlay it on the image. The actual polygon being uploaded doesn't have the wobbly features.
Now you're saying it's my fault for selecting a wonky outline. What's it gonna be, is the preview bad or the resulting polygons? (And the reviewer is bad for approving anything at all?)
> my idea was to show the potential, not providing a polished solution
I can appreciate that, but if you're aware of this then it shouldn't have a button that unauthenticated users can press to upload the result to the production database. OSM has testing infrastructure if you want to also demo that part (https://master.apis.dev.openstreetmap.org/ is a version I found on https://wiki.openstreetmap.org/wiki/API_v0.6)
I apologize. I read `it's uploading` and misunderstood like you were saying the tool itself was uploading things.
> is the preview bad or the resulting polygons? (And the reviewer is bad for approving anything at all?)
It can be one, the other, or both.
I was replying to a reference about a specific example in the blog post.
In that example, I see wobbly features due to rendering alongside the edges that make it look like the polygon is going to have dozens of nodes. Then, there is an over-simplification of the polygon around the top-right corner (which I didn't consider an error based on my criteria from reviewing manually created pools).
> And the reviewer is bad for approving anything at all?
I didn't say that. I was trying to assert that the UI/X can be improved to better show what will be uploaded.
> but if you're aware of this then it shouldn't have a button that unauthenticated users can press to upload the result to the production database
You are right. I was manually reviewing the profile created for the demo every day, but I didn't realize the impact/reach until I saw the first comments here. As soon as I read the first comment, I shut down the demo.
As I said in other comments, I will make alternative changes to the demo.
> if you want to also demo that part (https://master.apis.dev.openstreetmap.org/ is a version I found on https://wiki.openstreetmap.org/wiki/API_v0.6)
Thanks for the suggestion, I don't know why I didn't thought about that earlier.
I feel like a lot of the pushback here is an idea that OSM can grow from hand mapping; but as someone with 60k changesets over a decade... no amount of human volunteer enthusiasm is to the point that it can "solve" mapping at a global scale to the standards that make the map data overwhelmingly useful.
I feel we need a scalable framework for importing and maintaining data: ways to annotate the quality, sources, where to report bugs in the data source, and guidance to consumers. Ie if I want to query "businesses of type X" "mapped by humans within the last year", I can sort of do that with "check date".
But who knows how many of those attributes are accurate, or if the mapper who checked only checked one aspect (name/location)? Would it be better to ingest alltheplaces opening hours to maintain this data automatically, every month?
Would it be better as a data consumer if I could filter to only certain sources I trust? Or I could use data - even if the polygons aren't perfect or similar, even with known limitations like "poi inferred by AI".
I am working on such project
See https://community.openstreetmap.org/t/what-you-think-about-i...
https://www.openstreetmap.org/user/Mateusz%20Konieczny%20-%2...
https://codeberg.org/matkoniecz/list_how_openstreetmap_can_b...
Alltheplaces plays dangerously loose with (also) using resources clearly marked as copyrighted and protected with an API-key. As it is that project can serve as inspiration, but it is incompatible with OpenStreetMap.
I am currently working on project that would use ATP and I am vetting its spiders. So far I have not found any bad ones.
If you found one then knowing which one you mean would be highly useful!
(BTW, just marking something as copyrighted does not make it copyrighted)
https://community.openstreetmap.org/t/what-you-think-about-i...
There are probably more spiders configured to do this.
> (BTW, just marking something as copyrighted does not make it copyrighted)
For OpenStreetMap, it means that at the very least the Licensing Working Group should have a look. When you combine a copyright claim with directly using a third-party API with an API-key without clearance from the owner, doubly so. This was already pointed out to you in that thread.
Currently, only spiders which directly use the websites and domains of the shop chain (or its owner) are cleared for use.
Take heed of what Andy Townsend from OpenStreetMap's Data Working Group wrote:
> OSM has traditionally avoided situations where it could be legally challenged by people with more money to pay lawyers than we have, even if, in a fair and balanced process OSM might actually be in the right; for the simple reason being that any legal cost could far outweigh other costs of runnng the project.
You are a senior mapper in our project. You know this.
Because I wanted to know is there any other known case.
> For OpenStreetMap, it means that at the very least the Licensing Working Group should have a look.
and they were asked about first-party sources, that is why I am looking through ATP to check whether specific spiders are using only first party-sources or not
> Currently, only spiders which directly use the websites and domains of the shop chain (or its owner) are cleared for use.
AFAIK this is not accurate - for example if they would host their data at github pages website without custom domain it does not change things
So it might very well be automatic mapping, just saying that I know my armchair mapping can be pretty bad when I gonand check it OTG.
North Korea should be an interesting example, I guess we have no chance of anyone commenting on it but there's people who've been mapping it from actual satellite imagery, so very poor quality compared to the airplane-based georeferenced photographs we're used to in most countries. If armchair mapping makes bad maps, that will be the best example to have someone check out, but alas. (The visitor-accessible parts may not count as much because a mapper could have visited those)
Ola!
I am the author of the repo, worked in satellite projects for the Galician goverment some years ago.
You don't need an account to download the data from OSM (you do need to contribute back, which makes sense IMO). You don't need an account to download tiles from some publicly available sources (i.e. https://pnoa.ign.es/ in Spain) but I prefer to made the code work with MapBox and let them pay the infra (until they stop free offering). Happy to share a simple snippet to use a different tile provider.
Any public dataset (that I am aware) is not really meant to be frequently updated, at best you get a second version release a year later. A lot of public money (I know because have been payed a small portion of these budgets) is spent on building datasets that are used for a couple of research papers, uploaded to the web and then become outdated.
If I want to help updating them or correcting label mistakes, in most (all?) cases, there is no practical way for me to do it.
I believe that OpenStreetMap has the potential to be the best publicly available spatial "dataset" for (some specific) CV use cases.
If we focus on creating ways to contribute with quality data (the idea behind this small project), it will just keep getting a better dataset, that anyone can contribute to be up-to-date.
1.6. No Tracing, Deriving, or Extracting. Customer shall not trace or otherwise derive or extract content, data
and/or information from the Service Offerings except that Customer may use Studio or third-party
software to trace Mapbox Maps solely comprised of satellite imagery to produce derivative vector
datasets (i) for non-commercial purposes or (ii) for OpenStreetMap.
Please do not contribute anything ai-fantasized
If this can be somewhat consistent then it'll probably do better than the average OSM contributor. Something like segmenting houses, roads, bodies of water, comparing against current data and highlighting inconsistencies for correction would be a good start though.
Hi there, I agree that is a valuable usage of a model trained with OSM data.
I didn't have the time to release the code but I am/was doing exactly that to refine the training dataset. I take the trained model and run it against the ground truth from OSM. Any heavy mismatch between the two almost always result in an useful edit to be made in OSM.
let us reconsider this statement, please. An unexpected and powerful effect of the Openstreetmap project is iterated convergence on ground truth. No person is perfect in contribution, and few people are consistently terrible. Revision and updates, common vision towards accuracy, an appreciation of cooperative contributions.. have astounded the public and humbled critics repeatedly. Not because every key stroke and mouse click is perfect, but because iteration and plural sources have converged in a usable system of software and data.
AI inputs to Openstreetmap are not new, as noted in other comments. The path forward is bright, useful to humans and participatory in the Openstreetmap project.
I've got a Yolov8 model which does quite a good job of finding and segmenting out solar, but the edges are absolutely horrible and require an insane amount t of work to clean up. I've seen results from SAM2 that has been trained, and the results look massively better.
Wouldn't put these in OSM due to stereo6 comment about accuracy, but I could sure use them elsewhere.
Hi there! There is no SAM2 finetuning involved in the project. The segmentation data from OSM doesn't have enough quality to properly train a segmentation model from it.
What I am using here is a YOLO model for bounding box prediction. The bounding boxes from OSM are good enough for this. I then pass the each individual bounding box as a "prompt" to SAM2 for a segmentation of what is inside.
I also tried to pass the centroid of the box as a "prompt" for SAM but it gave worse results.
I have published a new release where any code to directly upload to OSM has been replaced with an export to OsmChange format.
I hope this is a step in the right direction, I will continue the discussion on the dedicated thread in the OSM forum.