Hacker News Reader

Comment by crazygringo | Hacker News Reader

Google is still providing driving directions over the collapsed bridge:

https://www.google.com/maps/dir/35.7822529,-81.2819178/3834+...

I got the location from LeifCarrotson on this comments page, who helpfully provided a link to the bridge in 3D Maps mode where you can clearly see the collapse:

https://www.google.com/maps/@35.7815814,-81.2828119,82m/data...

(I'm writing this comment as root-level rather than replying to LeifCarrotson, for higher visibility, so people can see this is still not a fixed problem.)

View on HN

And in case anyone is wondering, OpenStreetMap got it right! https://www.openstreetmap.org/search?query=35.781302%2C%20-8...

IANAL...but Google still serving up the wrong data - ~1 year after the fatal accident, and also after the lawsuit was filed against them - seems unlikely to score sympathy points in the jury box.

How has that road not been closed off?? That's the bigger news here.

It was in the Bing street view. I wonder how the signs got removed.

https://www.bing.com/maps?osid=c3224526-5ada-4a41-b8f7-e854f...

"Typically, barricades are in place to prevent drivers from crossing the bridge, North Carolina State Highway Patrol said. But the barricades had been removed after being vandalized and were missing at the time of Paxson’s wreck."

https://www.charlotteobserver.com/news/state/north-carolina/...

I can't find any reference to who removed the barricades, but that seems like the main criminally negligent action to me here. That's obviously going to get someone killed -- and it did.

Yeah this clearly seems like the fault of whoever is supposed to manage that road. Google maps also shows physical barriers there in May 2023, but different ones. Hard to say what it looked like at the time of the accident but if the road just disappeared I could absolutely see how that accident would happen with or without navigation.

https://www.google.com/maps/@35.7811991,-81.2834442,3a,16.5y...

Oh, this is crazy. So Google still provides directions across that bridge, and has data collection from May showing (qualitatively) the bridge is out. But when you go to StreetView, you can't hand-navigate across the bridge in the 2023 data, but if you snap back to the 2012 data, you can... And Google will let you navigate back and forth, then snap back to 2023.

Looks like the revision in 2023 didn't remove the road from the 2012 dataset, and Google's map resolution algorithm is chaining back to an older revision to find the connectivity that justifies the path. What a mess.

You can't go over the bridge in street view in 2023 because the street view car didn't drive over it (in this case because it can't). Clicking the map in street view chains to the nearest street view image.

They don't definitively use street view data to determine their route network for navigation; there's lots of roads that the street view car doesn't drive which are valid roads. For example forest resource roads are like that here in BC, they are allowed to be driven (when the roads are open) and will be navigated through by Google maps (regardless of if it's actually passable) but street view cars never go down them.

Correspondingly, you can use street view to move across roads that are not navigable at all -- for example backwards up one way streets or across intersections that are impassable by cars. The street view map isn't a good proxy for routability.

It's not, but in this case I suspect it's giving us a clue as to how the algorithm broke. Here's my imagined failure mode:

1. Google collected new data in 2023

2. That data didn't include a turn down the busted road (because why would they bother, they can see the bridge is out from up the hill)

3. Since there was no turn down the busted road, there was no new data collected for the busted road

4. They ran a fusion step to build a connectivity graph from snapshots of traversible world. Here in particular, that would have been necessary because there isn't much visibility on the road from the sky either, due to the tree line.

5. Their algorithm considered street-view-collected data to be highest-quality. Especially new street-view-collected data.

6. The connectivity graph went "Hm, I'm missing new data for this road segment... Oh, but I have some from 2012. I'll just assume we missed collection here and fuse the old data in to build a coherent map."

7. This fusion step killed all the "Hey this road doesn't exist" edits because it kicked out a new, "pristine" map from real-world measurements. In essence, the fusion step improperly updates the system's belief on recency of data so that the 2012 data is treated as fresh as the 2023 data and overrides older (2022 etc.) reports the bridge is out.

Steps 6 and 7 here are bugs, but it's a subtle enough bug that I can imagine the Maps team failing to catch it until disaster hits.

My point is that I do not think that google uses the street view car data for creating their road networks at all, because the data is effectively useless at that for a wide variety of reasons.

The first reason is that while GPS data has sufficient accuracy to, with markov models, correctly identify which road of an established road network you are on, it doesn't have sufficient accuracy to retrospectively create an accurate road network from raw GPS tracks. The second reason is that the route that the street view cars take would cover the road network for photographic purposes is not sufficient to enumerate the annotations that must be captured in a road network to facilitate navigation, especially once you consider that street view has not just driving navigation but also cycle and pedestrian navigation. The third reason is that the street view car will not drive most roads frequently enough to be sufficient for map maintenance and will not drive many roads at all ever, purely on a cost basis.

Google maps gets their road data from a series of partnerships with private companies and public entities. Those partners provide highly detailed, human maintained road network data which are integrated into an annotated directed graph which is then used for navigation. Street view cars fundamentally do not have enough information or accuracy to be an authoritative source for this, and it's far more likely that the street view cars get their driving routes from google maps' road data.

It's possible they consider some street view data in generating their map features. But again, there's a wide variety of reasons that street view might then not go down a road, so having google maps automatically remove a road from their dataset when the street view car doesn't go down it would result in far, far more numerous and problematic edge cases than this one (think roads disappearing because they had construction on the day google street view was in the area or the driver missing an area of town and having it disappear from the map). I have no insider knowledge and google is fairly quiet about their process but I would expect street view to be essentially a feature only built on top of the road network, and I would be extremely surprised if it functioned as the authoritative source for it.

If we look for a process improvement, I think what could have been a feasible organizational solution is -- since the google street view car has been down this path since the accident -- to encourage street view car drivers to submit corrections for the road network when it's incorrect and prioritize those. But it sounds like humans were already submitting this and nobody was doing anything and others have commented on how opaque the process can be so I've no idea what happened on that end.

> But it sounds like humans were already submitting this and nobody was doing anything

This is what leads me to believe that there was a data fusion error. The scenario that sounds likely is that they successfully flagged the street aa unnavigable, then something happened in the intervening time (such as a May 2023 updated data collection) that convinced the map it had newer real-world-facts to contradict the older human reports.

I agree that StreetView isn't sufficient on its own. But if Google is doing any kind of AI analysis of the road images, StreetView collection is the highest-accuracy data they could possibly draw from. This smells like "We trusted our algorithm for going from road images and car path metadata to map construction and it failed us."

> think roads disappearing because they had construction on the day google street view was in the area or the driver missing an area of town and having it disappear from the map

That's two reasons they can't rely on only StreetView collection; they'd have to have an algorithm for fusing StreetView data with other data sources (including older StreetView runs).

That's clearly an insufficient level of "closure" for a hazard like this. Those signs look like they could easily be removed by one or two people without any kind of specialized tools. An adequate barrier would be more like a few concrete blocks physically preventing anyone from driving past.

It looks like that's what they have there as of this year.

That's extremely common for roads in the "country" because there's not much available for the cost of installing jersey barriers across the road, etc.

But when the bridge is out such that you'd die, it probably calls for a more permanent blockages or at least sending out the bulldozer to put up some earthworks.

Why isn't the bridge fixed?

Article says it was a privately maintained bridge and none of the relevant property owners fixed it.

Another comment showed the Google link where the route was blocked which got me interested. And the solution is: the bridge is mapped as one-way street :) ! (try swapping the start+end and it won't pass the bridge)

Very confusing... any maybe the reason why it has been not properly fixed.

I don't see the issue. Was it fixed today?

Looks like it was indeed, in the past couple of hours. The entire road segment is now removed from Maps.

Nothing like a top comment on an HN front page story to get a company to take action, I guess.

Or maybe it was the coverage from multiple news organizations.

Reply on news.ycombinator.com