Here’s the number that matters most in this story, and it isn’t an accuracy score. It’s 1.7 percentage points.
A curb ramp at 1:12 has a slope of about 8.3%. At 1:10, it’s 10%. The first passes under the ADA and the second fails. Nobody can see that difference in a street photo, and plenty of AI tools now sold for accessibility audits work from street photos.
The rest of the rulebook is just as fussy. Cross slope can’t go past 1:48. The top of a curb ramp needs a level landing at least 36 inches long. Entrance ramps on buildings carry their own set of rules: 36 inches of clear width, handrails on both sides once a run climbs more than 6 inches, and 60-inch landings. (The fuller rundown in ADA requirements for wheelchair ramps covers handrails, edge protection, and landings, if you’re building or retrofitting one of those.)
A model that says “ramp detected” has answered the easiest question on the form.
Mapping Every Ramp Overnight Is the Easy Part
Cities still have to count their ramps. Under ADA Title II, any public entity responsible for streets has to put a curb ramp schedule in its transition plan. Lawsuits have pushed a few cities onto fixed timelines on top of that.
Counting used to be slow, sweaty work. Crews walked block after block with clipboards and digital levels, or volunteers clicked through street images one at a time. Project Sidewalk, out of the University of Washington, built a crowdsourcing tool for that clicking. Its labels now help train the models replacing it.
Several things made the switch possible at once. YOLO-family detectors shrank enough to train and run on a single consumer GPU, and street-level panoramas spread across most streets in big U.S. and European cities. So a pipeline can now sweep a whole district overnight and return a map of candidate ramps, missing ramps and obstacles.
That map is useful. It just proves less than it looks like it proves.
The Porto Study, Read Slowly
The best evidence so far comes from peer-reviewed research. A 2025 paper in IET Smart Cities, published by Ietresearch, describes a pipeline called ACAMAI built around a YOLOv8 detector. The researchers trained it on Project Sidewalk labels mixed with a local Porto dataset. Shown 2D sidewalk images, it caught curb ramps with 91% recall and 85% precision.
Things got messier with full 360° Google Street View panoramas. The team sliced each panorama into many overlapping views and scanned in tight 2° steps. Recall held up well: 90% for curb ramps, 93% for obstacles. But the model also kept spotting faraway ramps that nobody had labeled, and every one of those counted against precision.
| Task | Metric | Result | In plain terms |
| Curb ramps, 2D sidewalk images | Recall | 91% | About 1 in 11 real ramps slips past |
| Curb ramps, 2D sidewalk images | Precision | 85% | Roughly 15 of every 100 flags are false alarms |
| Curb ramps, 360° panoramas at 2° steps | Recall | 90% | Close to the 2D result, but compute-hungry |
| Obstacles, 360° panoramas at 2° steps | Recall | 93% | Catches most, with lots of duplicate hits |
Location is the quiet catch. The panorama tests ran on a locally annotated Porto set, and Porto’s historic cobblestone streets often have no curb ramp at all. A Phoenix subdivision looks nothing like that. Neither does a Boston brick sidewalk.
So What Happens With 20,000 Ramps?
Put the study’s rates on a mid-sized city and the abstractions turn into workload. With 20,000 ramps and 91% recall, around 1,800 never make the list. With 85% precision, someone also has to clear roughly 3,200 false flags.
The misses are worse, legally speaking. Each one is a liability nobody logged. A false flag only costs a reviewer a few seconds to dismiss. Then again, a few seconds times 3,200 adds up to days of somebody’s job.
Anyone who’s owned a smart doorbell knows this trade. AI security camera alerts still go off for darkness, rain, and the odd garden statue. A ramp detector has the same habit, just with sidewalks.
So retest any model on your own streets before you believe its totals.
The Slope Problem No Camera Angle Fixes
Back to those 1.7 points. A flat image can’t reliably measure them. Camera height bends the reading. So does lens distortion, and so does the grade of the street itself.
Then there’s everything a city sidewalk throws in the way. Parked cars cover the bottoms of ramps. Shadows wash out the truncated-dome warning surface. The imagery may also be years old, which means a ramp rebuilt last spring can still show up as the cracked one it replaced.
Some newer products bolt a vision-language model onto the detector, so findings come back as tidy English sentences. Reviewers like that, and it’s easy to see why. A note saying a ramp “appears compliant” still sits on a 2D picture with zero slope data behind it, though. Language models are good at convincing but wrong details, and polished prose is exactly what slides past a tired reviewer at 4 p.m.
Even the Scorekeepers Re-Score
Benchmarks move. RampNet, a curb ramp detector from the Project Sidewalk team, posted an erratum in July 2026 after re-scoring its evaluation. Reported recall on its gold-standard test set dropped from 93.5% to 87.3%. Precision ticked up a little.
Nobody was hiding anything. The team found a scoring issue and fixed it in the open, which is how research should work. It’s also a decent reminder that a headline percentage deserves a second look, especially one sitting in a sales deck instead of a public repository.
Before You Buy, Measure 50 Ramps Yourself
A pilot you control beats any brochure. Pull 50 to 100 ramps your crew has already measured by hand and line the tool’s output up against them.
Start with the misses. However many measured ramps the tool never found, that’s your actual recall on your actual streets.
Next, look at slope. Does the tool give a number at all? If it does, is the error band tight enough to tell 8.3% from 10%? Plenty won’t.
Check whether it flags top landings, cross slope, and detectable warnings, beyond just spotting a ramp. That last one matters: the Department of Transportation’s ADA standards require detectable warnings on curb ramps.
Two smaller questions catch a lot of weak products:
- What’s the capture date on each image? Toss anything older than your last construction cycle.
- How does it merge duplicates? Overlapping hits pad the counts and pad review time with them.
If a vendor won’t run that pilot or won’t break out error rates by category, file the product under mapping aids and budget for a real audit anyway.
What’s Working in 2026
The setups that hold up split the job in two. The model sweeps the network overnight and ranks likely problems. A person with a smart level then starts at the top of that list, and the crew’s slope readings become the record of compliance.
There’s a practical reason for that order. If a ramp ends up in a dispute, a logged measurement carries far more weight than a bounding box drawn on an old photo.
The open question is what happens once the sensors start supplying slope data on their own. Somebody will still have to decide whether to trust it.
Related: 5 Best Mobile Wallet Payout Providers in 2026: Fees, Speed & Security
