Aerial orthoimagery curb-ramp detection already exists in production (34k ramps at 1-inch GSD): treat it as an uncorrelated arbiter, not a competitor

[jonfroehlich]

The finding

Curb ramp detection from aerial orthoimagery already exists in production, not as a research idea. Douglas County (CO) trained an Esri GeoAI deep-learning model on ~700 sample images against one-inch-resolution county aerial photography and mapped 34,000+ curb ramps in 12 days, versus an estimated six months of human inspection. That inventory fed their ADA transition plan, and the county re-runs it automatically each time new imagery is flown (every two years). Esri has since shipped a Pedestrian Infrastructure Classification model in the ArcGIS Living Atlas.

On the research side, Tile2Net (Hosseini et al., Computers, Environment and Urban Systems, 2023) generates sidewalk, crosswalk and footpath polygons plus centerlines from sub-metre aerial tiles, reporting 90%+ recall in Boston and Cambridge. Notably it has no curb ramp class — its classes stop at sidewalk / crosswalk / footpath.

I raise this because I previously described aerial cross-checking of curb ramps as an open, unclaimed angle. That was wrong, and the correction matters for how we position RampNet.

Why the resolution gap is the whole story

The two results above are not the same technique at different scales — they are separated by roughly an order of magnitude in ground sampling distance, and curb ramps sit right at the boundary:

source	GSD	a ~1.5 m ramp spans
county aerial (Douglas County)	1 inch ≈ 2.5 cm	~60 px
typical sub-metre aerial (Tile2Net)	~30–100 cm	~2–5 px
best commercial satellite (Maxar-class)	~30 cm	~5 px

So: curb ramps are detectable from county-flown aerial and effectively invisible from satellite. That also explains why Tile2Net — designed for sub-metre tiles — maps sidewalks but not ramps. "Satellite imagery" is the wrong frame for this entirely; the right frame is high-resolution municipal orthoimagery.

What it means for RampNet — complement, not competitor

The aerial approach wins where we are weakest and loses where we are strong, and the split is clean:

Aerial is better at: no occlusion by parked cars, pedestrians or foliage; no viewing-angle or far-field degradation (which is exactly the annapolis framing problem); uniform GSD across a whole county; trivially complete coverage of every intersection, including ones no vehicle drove past.

Street-level is better at: global availability — GSV and Mapillary cover cities that will never commission one-inch aerial, which is the entire premise of RampNet's dataset generation; ramp condition and obstruction, which are invisible from directly overhead; slope and surface quality; and the missing curb ramp judgment, which needs to see the corner from a pedestrian's viewpoint and is half of Project Sidewalk's value.

That last point is the important one strategically. An aerial model answers "is there a ramp here?" It cannot answer "is this ramp usable?" or "should there be a ramp here and isn't?" — and the north-star framing for the AI labeler is find, tag and rate.

Why this is worth an issue rather than a footnote

1. Related-work obligation. Any RampNet-2.0 write-up that discusses ramp inventory without acknowledging that a county already did 34k ramps from aerial is incomplete, and a reviewer will find it.
2. It is a validation source we do not currently use. For any city that publishes high-resolution orthoimagery, an aerial pass is an independent detector whose errors are uncorrelated with ours — no shared occlusion, no shared viewpoint, no shared imagery vintage. That is a far stronger cross-check than another street-level model, and it is a candidate arbiter for the same role Experiment: hybrid RampNet + open VLM (Qwen3-VL) — ensemble / cascade / distillation #35 gives the crop classifier.
3. It bounds the multi-view work in Multi-capture / multi-view fusion for per-ramp recall: correspondence, agreement-as-arbiter, 3D (RampNet 2.0) #48. If the goal is a ramp's map position, aerial gives near-ground-truth positions wherever the imagery exists — which is both a way to evaluate triangulation accuracy without field survey, and a reason not to over-engineer triangulation for cities that have the aerial.
4. It interacts with Extend RampNet output from points to boxes/polygons (point-prompted SAM2; CenterNet-style size head; segmentation) #83. Cross-checking against aerial requires extent, not points — a ramp polygon is what you can actually align to an orthophoto. Extend RampNet output from points to boxes/polygons (point-prompted SAM2; CenterNet-style size head; segmentation) #83's path 1 (point-prompted SAM2) is the cheapest route to that.

Possible first step (cheap)

Pick one benchmark city that publishes high-resolution orthoimagery, pull the tiles for the panos we already have GT on, and measure agreement between RampNet's triangulated positions and what is visible from above. That gives, in one pass: an independent recall estimate, a position-accuracy number for #48 that needs no field survey, and a concrete read on whether aerial is worth wiring in as an arbiter.

Worth checking first whether any of richmond / bend / clovis / morgantown / annapolis publishes orthoimagery at better than ~15 cm — the approach is only meaningful above the resolution threshold above.

Related

• Multi-capture / multi-view fusion for per-ramp recall: correspondence, agreement-as-arbiter, 3D (RampNet 2.0) #48 — multi-view fusion; aerial provides position ground truth for evaluating triangulation
• Extend RampNet output from points to boxes/polygons (point-prompted SAM2; CenterNet-style size head; segmentation) #83 — points → boxes/polygons; extent is a prerequisite for aligning to orthoimagery
• Experiment: hybrid RampNet + open VLM (Qwen3-VL) — ensemble / cascade / distillation #35 — hybrid cascades; aerial is an arbiter with genuinely uncorrelated errors
• Second rater for Budapest District V: the rubric is now confounded with the benchmark's only ranking inversion #74 — the budapest rubric question; non-US cities are exactly where aerial will not exist

Sources: Esri / Douglas County GeoAI curb ramp inventory · Tile2Net paper · Tile2Net code · StateScoop on municipal GIS curb-ramp work

🤖 Generated with Claude Code (claude-opus-5[1m])

[jonfroehlich]

Aerial as a source, not just an arbiter — plus the correspondence math and a GSD floor

Follow-up from a working session on #59. This issue frames aerial as an uncorrelated arbiter, which
is right, but three things push it further: the correspondence problem is much easier than the
street-to-street case, the resolution threshold is now pinned to specific published tiers, and
aerial may attack the sourcing bottleneck rather than only validating against it.

1. Correspondence is asymmetric, and lat/lng does most of the work

Aerial ortho is georeferenced by construction — every pixel already carries a lat/lng, so
detections come out as map coordinates with no correspondence step at all. All the difficulty sits
on the street side: back-projecting a RampNet pixel detection to lat/lng needs pano pose plus
distance, and docs/detection_recall_analysis.md puts the flat-ground estimate within
6.5–8.5% of DA3 metric depth (Spearman 0.95 GSV / 0.81 Mapillary).

ramp at	~8% distance error	vs a 1.2 m ramp
10 m	0.8 m	inside one ramp width
18 m	1.4 m	~one ramp width
30 m	2.4 m	2× ramp width
40 m	3.2 m	2.7× ramp width

So the match tolerance decides feasibility:

• Corner-level ("is there a ramp at this corner?") tolerates ~5 m and works across the whole
  range. This is the fusion case, and it is fine.
• Ramp-level ("which of these two ramps?") needs sub-1.5 m and degrades past ~15–20 m — i.e.
  exactly paterson's paired-TSI problem. Though if the aerial GSD resolves both surfaces, aerial
  answers that on its own and the match is unnecessary.

Note the awkward part: correspondence is hardest precisely where aerial is most valuable (far
field). Corner-level matching is the way around it.

2. Why it is complementary, in numbers measured this week

A new miss decomposition across the seven US splits (scripts/analysis/miss_decomposition.py, #59)
finds 57.8% of RampNet's misses are beyond 18 m — 0.120 recall points of a 0.207 total gap.
That population is pixel-starved, and E1 showed it is not inherited from the labels.

Aerial has uniform GSD, so it has no far-field cliff at all. Our biggest miss population is an
artifact of street-view perspective geometry that simply does not exist overhead. That is a stronger
statement of complementarity than "different failure modes".

3. The GSD floor, and which published tiers clear it

GSD	tier	1.2–1.5 m ramp spans
60 cm	NAIP standard (since 2018)	2–2.5 px	❌
30 cm	NAIP state upgrade; USGS HRO 1-foot	4–5 px	❌
15 cm	USGS HRO 6-inch; some city	8–10 px	⚠️
7.6 cm	city 3-inch	16–20 px	✅
2.5 cm	county 1-inch (Douglas County)	48–60 px	✅

The free federal tier does not clear it. NAIP is standardized at 60 cm with an optional 30 cm
upgrade (~half the states at 30 cm in the 2025 acquisition) and is flown leaf-on, the worst case
for canopy occlusion. USGS HRO is better — "1 m or better", many urban areas at 2-foot or finer,
some at 1-foot, and leaf-off — but still lands at or under the threshold for a ramp.

Restating this issue's point in terms of the frame: it is not satellite, and it is not federal
either. It is municipal/county-flown ortho.

4. AccessParkCV is in-house precedent, with one caveat that matters

AccessParkCV (Hwang, Li, Kang, Hosseini,
Froehlich; ASSETS 2025) detects disability parking and access aisles from aerial orthoimagery at
micro-F1 0.89, using 7.62 cm/pixel for Seattle, Washington DC and Spring Hill (15 cm for the
Massachusetts evaluation), pulled through Tile2Net's orthoimage tile system from
city-published sources.

The caveat: parking spaces are ~2.4–5 m wide and access aisles ~1.5 m, so at 7.6 cm those span
30–65 px. A curb ramp at the same GSD spans 16–20 px. AccessParkCV was working with objects roughly
3× larger in linear extent — effectively 2.5 cm-equivalent imagery for its object class. Its
success does not transfer directly, and the open question is where the floor is for a ramp.
Douglas County proves 60 px works; nobody has established the minimum.

5. The bigger claim: aerial may bypass the sourcing wall, not just validate it

Session work on #59 (see docs/curb_ramp_data_sourcing.md, PR #93) mapped how hard government
inventory sourcing actually is: roughly 10% of large US cities publish curb-ramp data at all; only
3 of the 8 cities the paper assessed rated "Good" for location precision, and all three are
already in training; Chicago has 137,000 ramps and publishes none of them; Phoenix's rich
inventory sits behind a city login.

RampNet's Stage 1 premise is government GPS points + street imagery → training labels. Aerial
detection can manufacture the GPS points. For any city with sufficient ortho, that routes around
the open-data bottleneck entirely.

There is already a live proof of this in our own candidate list, which we found while hunting for
government inventories and only noticed at the schema check: Denver's 72,770-ramp layer is
literally "delineated from 2022 aerial imagery." It also came out as the best temporally aligned
large candidate in the #59 assessment. Douglas County and Denver are two production instances of
this exact pipeline.

6. A sharper first experiment than "find a city with good ortho"

Seattle and Washington DC both publish 7.6 cm ortho (AccessParkCV used both), and the lab
already has tooling that ingests that tile format.

The irony is useful: on the inventory side Seattle is rated Poor (paper Tab. 1) and DC has no
install-date field at all with YEAR_INSPECTED = 2016 on all 34,859 records — both damaged as
government data sources. But both publish excellent imagery. Aerial detection could produce for
Seattle and DC exactly what their own published inventories fail to deliver: accurate, current ramp
locations.

So the first step becomes concrete rather than exploratory — we know the cities, the resolution
clears the bar, the tile tooling exists, and there is benchmark GT to score against:

1. Pull 7.6 cm tiles for a bounded area with existing GT.
2. Hand-label a small ramp set to establish whether 16–20 px is workable at all — the floor question
   above, and the cheapest thing that could kill the idea.
3. If it holds, score aerial detections against benchmark GT at corner-level tolerance, which
   §1 says is the achievable match.

That yields, in one pass: an independent recall estimate, a position-accuracy number for #48 needing
no field survey, and a read on whether aerial can serve as a source rather than only an arbiter.

Unchanged from this issue's original framing: aerial cannot see slope, surface condition, or
obstruction, and cannot make the "should there be a ramp here and isn't" judgment from overhead. #86's
measurement-and-condition scope still needs street-level imagery.

🤖 Generated with Claude Code (claude-opus-5[1m])