Get Your POI Out of My Parking Lot

The open sourcing of massive “Point of Interest” (POI) datasets from Overture and Foursquare has been a massive boon. A lot of the discourse around these data sets has been how to leverage various signals to improve the currency and depth of the data. Is the POI still in business? What are the operating hours? What is the website URL? One key area that tends to get overlooked is the accuracy of the geographic coordinates assigned to the POI.

The Pitfalls of Geocoding

The traditional backbone of POI datasets are business listings that contain an address for the place. These addresses are then processed by a geocoder to determine the latitude and longitude coordinates for the business. Geocoders are a wonderful, but inherently imprecise and blunt instrument for determining the location of a POI. During our days at GeoCommons we built an open source geocoder with support from a FGDC grant. A few takeaway that still hold true:

1. Addresses Aren’t Exact

When we use street addresses, we assume they’re precise. But addresses were designed to help people, not computers, find homes and businesses. Sometimes one address refers to an entire property, which can be quite big. A geocoder might pick a random spot on the property rather than exactly where the main building or entrance is.

2. Missing or Old Information

Geocoders rely on databases that aren’t always current. If a building is new, renovated, or moved, the geocoder might not know about it yet. This can lead to the marker appearing down the road or in an empty space nearby.

3. Guessing Along the Street

Sometimes, geocoders don’t have exact coordinates for each address, so they guess by spreading addresses evenly along a street. If your building doesn’t fit neatly into this pattern, the pin could end up in the wrong spot.

4. Shared Addresses

Large buildings like malls, apartments, or office complexes often have a single address but multiple entrances or businesses inside. A geocoder might place the marker at the main entrance or driveway, which may not match exactly where you’re headed.

5. Maps Can Be Slightly Off

Even when addresses and locations are correct, sometimes maps themselves aren’t perfect. Building outlines on maps might not line up exactly with real-world locations, causing confusion.

Here is how these factors can wreak havoc with POI placements on a map. In the map below the shopping center is set back from the main road with an access road. The address for the POIs, though, is the main road, which leads to a variety of POI placements.

Overture is not alone in this challenge. Foursquare has similar problems with their recently open sourced data.

There is a discontinuity across sources on both the placements of POI’s and which POI’s are present at the locations. Often business listings are acquired from multiple sources. Some of these have lat/lon coordinates already, but the provider used a variety of different geocoders with varying accuracy. This can cause a mix of results or duplicate entries seen above.

Google Supremacy?

Google has arguably done the most to address these concern in their decades long mission building their map products. This has led to a pretty amazing conflation of POIs to building footprints, which are then aligned with street view imagery. Take this example of a shopping center, shown in Street View, to illustrate how well aligned the POIs are with reality.

While Google really nails this challenge in Street View the results in Google Maps are a little less spot on:

While Rocky Mountain Tap & Garden and Renew Movement are correct in Street View they are inaccurate in Google Maps. Even for Google the POI problem is really challenging, and your POI can easily end up in the parking lot.

How Accurate Do POIs Need to Be?

For most 2D mapping applications it is arguable you don’t need this level of accuracy. Lots of location based apps have been massively successful without building or entrance level accuracy. Once a routing algorithm gets your “little blue dot” gets close enough to the POI our human abilities to orient and recognize landmarks take over. The last meter orientation gets us to where we need to be. What Google knows is that for augmented reality, and later XR smart glasses, their POIs need to line up with the real world. The impressive thing is this work started before 2020, back when AR style overlays were added to Street View.

How Can Open Data Places Compete?

Keeping pace with a juggernaut like Google (or Apple for that matter) is a daunting task. Fortunately consortiums like Overture and experienced players like FourSquare have added a lot of content to the open source community. Given these wonderful tailwinds we thought there could be an opportunity to contribute. Specifically, can we generate building or entrance level accuracy for the location of open POI databases?

POI Relocalization

The technique we wanted to try out was method called relocalization. This approach inverts the visual positioning problem of using images to identify the location of a user. Instead we will use multiple images to determine the location of a POI. If we know where the photos were taken from and which way the camera was facing, we can trace lines from where each photo was taken back toward the thing in the picture. Where those lines cross? That’s our location.

The use case we wanted to simulate was a future smart glasses user asking the question “Tell about the place I’m looking at”. The image collected by the smart glasses would then be an intentional photo of a POI we could use for relocalization. To simulate this use case we built a little mobile app that detects what Overture POIs that are close to you, then you select it and take an image of the POI.

We then used the app to take images of multiple POIs throughout downtown Louisville, CO. Once we’d collected the data we ran the relocalization algorithm to see if we could improve the location of the POI’s in the Overture database. Here is a map of our results where the green marker is the original location in Overture and the blue marker is the relocalized position.

We can then calculate the distance between the Overture marker and the relocalized marker as a blue line. Creating a frequency distribution of the length of the blue lines can show how well we improved the position in meters.

Building on this approach the ideal next step is to align the POI with the building footprint it is located in. Fortunately, Overture has a wonderful tool for facilitating this type of data conflation. GERS ID’s are unique identifiers for each feature in the Overture database. This allows us to then conflate a POI’s GERS identifier with the specific building it is located in. With this relationship between POIs and buildings created we could then begin breaking out entrances for the building GERS identifiers associated with the appropriate POI.

Overall we were pleased with these results, but there is a challenge with scaling out this approach. Some day there will be a huge number of users with smart glasses asking about POIs, but today is not that day. To scale out this approach we will likely need serendipitously collected images of places or a gamification mechanism to incentivize intention POI photo collection. In our next blog post we will discuss how crowdsourced street view imagery sources, like Mapillary, could be used to correct POI locations at scale.