I can find courts in Oklahoma by entering "Oklahoma" in the search box and selecting Oklahoma from the dropdown menu. Are you not getting the dropdown? It sometimes takes a few seconds for it to appear after you enter your search term.
I zoomed into a handful of random spots, and quite a few (maybe 30%) do not appear to show basketball courts at all. Many seem to be just rooftops or construction sites.
Unlikely, I zoomed in on one that had a pool with round edges within a square concrete area that had the rough geometry of a basketball court and I imagine a image recognition engine being confused by that.
The labels seem to be taken from OpenStreetMap, so blame imperfect labelling there.
I guess taking labels from OSM is akin to taking your facts from Wikipedia (except that Wikipedia actually tends to be pretty well curated these days).
In fairness to OSM volunteers, a lot of Wikipedia's subject space is historical or otherwise relatively stable, while everywhere OSM maps is in danger of being changed in the real world. Just comprehensively mapping one small town takes many hours and by the time you work your way through region much of what you mapped will be different IRL. Aerial photography only updates periodically, which helps in a sense but also hurts because anything you map with that will most likely be outdated from the start.
Doing some research on schools, parks, churches, etc. where courts commonly are implies the number of courts here may be way too low. I assume this was a model trained by manually curated examples, so I wonder if a class of examples was missing from the training data.
I randomly thought of one in my city and went to look for it. It was missing, at 28.5860615, -81.3010721 or at least the court doesn't come up if you type the ZIP code there into the search box.
The labels come from OSM so it was probably done by somebody manually looking over Bing's aerial photography and manually tracing what they can see. The iD editor uses Bing not Google.
there's ~0% chance there's only 50K outdoor basketball courts in the US. Other comments observe a ~30% false positive rate. So this is probably "a yeoman's ML pass at a lot of map tiles I scraped" more than "satellite imagery of every outdoor basketball court"
I guess as an example of using AI image processing on satellite imagery. E.g. instead of basketball courts you could let it automatically find military targets...
Are there any technical details available on how these images were captured. I'm assuming some form of Vision AI (YOLO?) was leveraged. What source data was used, and how was it processed?
I'm interested in doing something similar for tennis courts.
45 comments
[ 1.3 ms ] story [ 95.7 ms ] thread0: https://maps.app.goo.gl/xvC9s6dVo264JFaKA?g_st=ic
1: https://maps.app.goo.gl/4aTmEyxG1Fa2GFRKA?g_st=ic
I guess taking labels from OSM is akin to taking your facts from Wikipedia (except that Wikipedia actually tends to be pretty well curated these days).
I'm interested in doing something similar for tennis courts.
This could be done with OpenStreetMap metadata.
--
EDIT: I was right. In the "About" dialog:
> Court Locations via Open Street Map. Satellite imagery via Google.
https://jk-lee.com/projects/the-big-atlas-of-la-pools/