Rapid is a free, open-source, web-based editor for OpenStreetMap. In the past the focus was on conflating AI-generated datasets with OpenStreetMap data, but the future for this editor is conflating authoritative datasets with OpenStreetMap.

Humans are in the loop — people reviewing authoritative datasets and adding them to OpenStreetMap with a few clicks.

So you might be wondering, what is authoritative data? And perhaps it doesn’t even matter what authoritative means — maybe the most important thing is whether it is correct.

If you are interested in OpenStreetMap, you might enjoy this episode: OpenStreetMap Is A Community Of Communities, which is a great introduction to OpenStreetMap as a project but also explains some of the commercial interest in updating the map, which adds a lot of context to Rapid and its development.

Resources mentioned

• Humanitarian OpenStreetMap Team — well worth checking out.
• Segment Anything (SAM) — can segment objects by clicking or interactively selecting points to include or exclude.
• Mapillary — a platform that provides street-level imagery and map data from all over the world.

In Conversation

Becoming a “Dishonest Geographer”

Daniel: Ben, welcome to the podcast. You’re a software engineer at Meta in the Maps Geo team. How did you end up there?

Ben: I’m a web developer at Meta, and you could say I came by geospatial a bit dishonestly. I joined the group without fully realizing what they did — which was a lot of awesome work, especially in the open-source community. After I joined, I got a chance to start working on a tool called Rapid, and I’ve been working on it ever since, for almost four years now. Most of my previous software career was spent at cloud network storage and middleware places, where I didn’t get a lot of chance to work directly with an open-source community or with front-end tech. Rapid lets me do both, so it’s been a wonderful job.

What the Rapid Editor Is

Daniel: So what is Rapid?

Ben: Rapid is a free, easy-to-use, web-based, open-source map editor for OpenStreetMap. Any of your listeners familiar with OpenStreetMap who want to edit it can point any browser to rapideditor.org, click the “start mapping” button, and they’re off to the races. If you search for “Map with AI,” that’s also Rapid — it’s a soon-to-be-retired domain name we used when we first launched, and we’re retiring it because Rapid is now a lot more than just mapping with AI.

Mapping with AI-Generated Data

Daniel: What do you mean when you say “map with AI”?

Ben: We don’t mean the tools themselves use AI — we’re talking about AI-generated data. These are road and building shapes generated from orthographic satellite imagery using neural nets and deep learning. Meta created a global roads dataset, and Microsoft created a buildings dataset. My very first exposure to the OpenStreetMap community was being sent to a Youth Mappers mapathon in Washington DC, where people asked if there was a way to make adding buildings and roads easier — the state of the art was painstakingly tracing shapes on top of satellite imagery. That’s the jumpstart for why Meta and Microsoft looked at using AI.

Daniel: Why not just dump that data straight into OpenStreetMap where there are gaps?

Ben: “Human in the loop” is a phrase near and dear to our heart. We want to make absolutely sure no data gets into the map that isn’t high accuracy. We have incorrect data in both datasets — we may have predicted a road where there’s actually a stream, or a building that’s actually a parking lot. So we want humans to review every shape. When we launched back in 2019, we were aware these AI changes might be controversial, which is why we spent so much time making sure you couldn’t add things directly — we want editors to think about each shape, not blindly click. A community survey later showed OpenStreetMap overwhelmingly voted that this is just another tool.

Authoritative Data and the Esri Integration

Daniel: Can AI be authoritative?

Ben: The Geographer of the US covered this in his opening remarks at State of the Map US recently — he didn’t love the term “authoritative” because it could muddy the waters. He just wanted to make the point: I don’t care if your data is authoritative, I care if it’s accurate. I don’t care who commissioned its creation. So I wouldn’t consider AI data authoritative, precisely because we’re asking editors to make that case-by-case decision about whether it’s good enough. Maybe once a human has looked at it and confirmed it’s correct, then it becomes authoritative.

Daniel: How is authoritative data getting into Rapid?

Ben: Users can open an ArcGIS dataset browser — an ArcGIS-hosted OpenStreetMap community API. These are datasets looked at by Esri’s QA folks and licensed to be ingested into OpenStreetMap. Esri works with local governments — a county seat or a township — and takes addresses and building shapes into the integration. The advantage is they’re authoritative, high quality, and very accurate; their only disadvantage is they’re hyper-local — a single ZIP code or county at a time, rather than global like the AI datasets. The conflation process is largely the same. Rapid has a back-end service running conflation behind the scenes, fetching data from Esri’s API and OpenStreetMap so you only see the buildings that are missing.

Mapillary and Street-Level Predictions

Daniel: How does Mapillary play into this?

Ben: Mapillary is a crowdsourced street-view imagery company — one of our sister organizations in Maps Geo. You sign up, capture images with your phone or a 360-degree camera mounted on your car, and upload them. Mapillary runs post-processing — preserving privacy by blurring faces, and predicting features from the images. They’ll analyze the imagery and place points on the map for utility poles, street lights, fire hydrants, bike parking. We have an integration in Rapid for viewing Mapillary imagery, and soon people will be able to use Mapillary predictions — objects that would be almost impossible to detect from orthographic satellite imagery taken from space.

Humanitarian Mapping and Tasking Manager

Daniel: Could you talk about the work of the Humanitarian OpenStreetMap Team?

Ben: The Humanitarian OpenStreetMap Team has a software bundle called Tasking Manager. If you have a bunch of editors who want to help after an earthquake, it’s not sufficient to just send them to a country and tell them to go edit — you need structure. Tasking Manager subdivides the area into a few hundred vaguely square sections for people to map individually. In February 2023 our daily active users for Rapid skyrocketed to something like 20 or 30 times normal, because we’d worked with HOT to get Rapid published as one of their editors, and project managers working on the Turkey earthquake response were using it.

Daniel: Do you build anything into Rapid to catch vandalism or low-quality edits?

Ben: Our first line of defense is probably OpenStreetMap’s biggest feature — the community, which polices malicious edits very effectively. That said, Rapid has a built-in limiting factor: once you’ve added 50 AI or authoritative bits of data, you’re asked to commit what you’ve done. That cuts down on enormous change sets where people click as fast as they can. We also tag every change set with a marker saying which dataset was used, so if there’s a problem it can be traced.

Rapid V2 and Building for the Community

Daniel: You’ve got a technical challenge, a design challenge, and a cultural challenge — the community was around long before Rapid. Which is hardest?

Ben: The community is never far from our thoughts. We constantly weigh how impactful our work is — if we implement a feature, how does it improve the mapping experience for enough people? That’s one of the guiding principles that led to Rapid V2, which was fundamentally rewritten from the ground up to be fast. Rapid V1 was slowing to a crawl on even medium-complexity scenes — your laptop fans would spin up and you’d get one or two frames per second. That was a problem affecting 100% of our users and only going to get worse, so we spent essentially all of 2022 rewriting things to make Rapid much faster.

Daniel: Isn’t there a tension — won’t the community just ask for faster horses when you could give them a car?

Ben: For sure, and part of it is just engineering time — the developer team for Rapid is two people, myself and my colleague Bryan Housel, who’s been a fixture in the community for years. We don’t have enough hours in the day to accommodate every feature request, and sometimes we have to say yes, that’s a great idea, but it’s either not impactful enough for our total user base or too technically difficult. It never feels good to make those judgment calls, but that’s software engineering.

The Future: Map Gardening

Daniel: Give us a peek into the future of Rapid.

Ben: Now that we have a much faster renderer, we’re thinking about ways to support not just adding data to the map. It’s easy to look at a map tile and tell whether it’s well mapped, but much harder to tell whether it’s up to date. So we’re thinking about editing workflows that promote “map gardening” — taking data that already exists, recognizing that a city is well mapped but maybe a shop has closed or sidewalks have been added. Those are much harder things to surface in a way that’s understandable to new mappers. I’d also love to give people more say in what’s rendered — someone into bike infrastructure could open a preferences dialog and say “just show me bike parking and bike lanes, styled this way.” As a developer I can’t anticipate every use case, but I can build tooling that lets people support their own.

Daniel: If there’s one thing you wish people understood about Rapid, what is it?

Ben: The question we get most at conferences is “why didn’t I hear about this — how long has Rapid existed?” So it sounds like we have a marketing problem. We hope everybody uses Rapid as their day-to-day OpenStreetMap editor. We exist, we’re here, we’re very willing to entertain feature suggestions, and it’s open source on GitHub. The short answer to “why should I use this” is: it’s much faster than it used to be — frankly a joy to use — and we have tons of data for you to use in your editing workflows.