In this week’s episode, I’m thrilled to welcome back Ariel Seidman, founder of HiveMapper. Ariel was my very first podcast guest back in 2019, and HiveMapper has come a long way since then!
We explore how HiveMapper has evolved from a drone-based mapping system to a cutting-edge platform collecting street-level data at a global scale. Ariel shares the challenges of scaling large-scale mapping efforts, the pivot to building their own hardware, and the role of blockchain-based incentives in driving adoption.
Here are just a few topics we cover:
- Why HiveMapper shifted focus from drones to street-level mapping.
- The power of combining hardware and software to solve mapping challenges.
- How HiveMapper has already mapped 28% of the global road network.
- The revolutionary edge computing and data filtering techniques driving efficiency.
- What it takes to compete with industry giants like Google Maps.
Whether you’re fascinated by the intersection of geospatial technology and innovation or looking for insights into scaling impactful startups, this episode is packed with value.
Let me know your thoughts or hit reply if you’d like to discuss the episode!
Connect with Ariel here https://www.linkedin.com/in/aseidman/
PS
I have just finished creating a web-based tool that lets you explore and download OpenStreetMap data, It is a bit different from other tools and I would appreciate some feedback.
https://mapscaping.com/openstreetmap-category-viewer/
In Conversation
From Yahoo Maps to HiveMapper
Daniel: You were my very first podcast guest, back in 2019. For anyone new, would you introduce yourself?
Ariel: I’ve been in and around mapping and large-scale data collection systems since my first job out of school at Yahoo, around 2004–2005. I sat in the Search Group — roughly 25% of all queries had some local geo dimension — and ultimately ran Yahoo Maps, so I had an early, excellent seat to watch how Google Maps was getting built. Quite frankly, they kicked our ass, because they committed huge amounts of capital and technology to collecting data from the physical world at scale, and Yahoo wasn’t willing to. Fast-forward to today, Google is massive, they acquired Waze, and there are countries entirely reliant on Google Maps — if it stopped updating or got pulled, that country’s GDP would actually suffer. There are maybe one, two, or three large-scale global maps at most, and in some countries only one.
Why HiveMapper Left Drones Behind
Daniel: Last time we talked, HiveMapper was focused on drones. What is it now?
Ariel: The vision has always been the same — how do you build a new large-scale global map from the ground up. We started with drones because they have a unique perspective: flying 100 to 200 feet up, you get a lot of both the aerial view and the street-level view — almost two for one. We built a whole software stack behind that. But two things hurt our ability to scale. One, battery technology never dramatically improved — a commercial DJI drone flew maybe 35 to 40 minutes, then you’d have to land it and swap batteries, which isn’t passive at all. Two, the legal side: every little city had its own drone regulations, and there were no federal laws coming any time soon. So we put drones aside, went down to street level, and iterated through a bunch of approaches until we found the one that’s scaling today.
Building Their Own Hardware — and Crypto Incentives
Daniel: What worked, and why?
Ariel: We tried iPhones and Android devices, third-party dash cams, even expensive external 3D camera systems — and ultimately decided to build our own device. Phones aren’t passive at all: you mount your one and only phone on the windshield, it overheats in summer, and every trip you have to open the app and hit play — that’s a tremendous amount of friction, so people just stop. And smartphone positioning is bad for mapping, off by anywhere from 3 to 15 metres. Dash cams at least start recording automatically, which is a big win, but their positioning is atrocious for mapping and you can’t access the onboard compute. So we bit the bullet and built our own hardware. The first device was ugly, but it solved the problems — it was entirely passive, high-precision, and we had total control. We deployed about 60 units across LA, Lagos, and Manila, and Manila lit up incredibly quickly. We knew there was something powerful there.
Daniel: And where did the crypto incentives come in?
Ariel: We were paying mappers small amounts of cash, but sending cash to places like Lagos or the Philippines is really complicated. People in the Philippines asked us to just send crypto — through PayPal they’d get paid Sunday but not see the cash until Thursday or Friday; with crypto they could cash out within 20 or 30 minutes and pay for groceries. We also saw another project on the blockchain incentivising people to deploy wireless infrastructure, and that bootstrapping model was similar to what we needed — getting people to deploy camera systems in vehicles all over the world. We launched with our own token, called Honey. We raised $21–22 million, built the next versions of the device, and one year after launch we’d mapped 10% of the global road network — two years after launch, about 28%. But that incentive mechanism doesn’t work unless the technology and product are entirely passive; it all has to fit together.
How the Sensor Works
Daniel: Let’s talk about how the sensor works.
Ariel: People ask why you can’t just use an iPhone — it’s a very good camera, but it’s designed for selfies, and it does a lot of processing on the imagery that isn’t great for mapping. Our device has three cameras. There’s a main wide-angle, high-resolution imager that, combined with a compute module, detects objects — that’s a speed limit sign, a stop sign, a turn restriction, this many lanes. But detecting an object doesn’t tell you where it is. So we also have stereo depth cameras with a wide baseline, far apart from one another — they work like your two eyes, giving spatial awareness — which lets us position objects very precisely, down to 25 to 50 centimetres. That depends on the device knowing where it is: we use a high-end GNSS module, an antenna five to ten times larger than a smartphone’s, plus IMU, compass, and visual odometry. We can position the device itself to roughly 50-centimetre accuracy. For a stop sign 3 to 20 metres away we consistently get it down to 50 centimetres; beyond that it improves the more people see it.
Route Intelligence vs. Location Intelligence
Daniel: What kinds of objects are important to you?
Ariel: Two broad categories. The first is route intelligence — everything that affects navigation: how many lanes there are, lane definitions, speed limit signs, stop signs, turn restrictions, street names, highway exit signs, toll prices. That’s the highest-priority category. The second is location intelligence — understanding something about a specific location: that the coffee shop over there is called Phil’s Coffee and it’s still open. If it has umbrellas out and people walking in and out, we can determine the business is still open; if there’s road construction next to a lot of Phil’s Coffees, maybe that retail business starts to suffer. It’s an interesting market, but definitely secondary to routing.
Edge Computing and Detecting Change
Daniel: You built your own device partly to process things on the edge. Does that mean you’re heavily filtering data?
Ariel: Day zero we uploaded all the imagery and processed it on AWS — that’s a lot of imagery. Now we do the mapping and detections on the Bee device itself, and only send up the relevant imagery and detections. If we see a 35 mph speed limit sign and there’s no change the next day, we just note “saw it, same” rather than re-uploading. You have to be efficient, because LTE and Wi-Fi aren’t limitless and AWS bandwidth and storage costs add up at scale. The firmware is always being updated — if we want to start detecting billboard signs, that’s a straightforward update — and because we see a given intersection in Phoenix maybe 180 to 300 times a year, backwards compatibility isn’t really an issue. The device develops a hypothesis about a location — this intersection should have four stop signs — and once it confirms that, it shifts into change-detection mode, which is more compute-efficient.
Competing with Google: Measurement, Not Inference
Daniel: Why would I choose you over Google?
Ariel: Two parts. One is basic map maintenance — the world is constantly changing: speed limits, turn restrictions, new stop signs, a lane going from two to three. Google and Waze mostly infer what’s going on from their motion data; Google Street View updates at best every 14 to 18 months, so day to day Google is effectively blind to those changes, or relies on subjective, noisy human reports. That’s one way we’re fundamentally different.
Daniel: So you’re measurement, as opposed to inference.
Ariel: That’s an excellent way of saying it. And don’t just trust me — most of the major mapping companies are using HiveMapper data. The approach makes sense because it’s high-precision and has the imagery attached: you don’t have to trust our object detection, you can look at the associated imagery and verify for yourself that a turn-restriction sign has been introduced. The other way we’re different is the real-time eyes — we can understand what’s actually going on at a location, distinguishing a little fender-bender that clears in 30 minutes from a four-car crash with ambulances that won’t. We have full density in Singapore, Amsterdam, Taipei, and parts of LA. To fully cover LA — about 100,000 road kilometres — you’d need three to four thousand active daily devices, ideally commercial drivers who are on the road a lot. For the top 10 to 15 US metros, probably by the end of next year.
Lessons: Persistence and Distribution
Daniel: What’s been the biggest takeaway from HiveMapper?
Ariel: Persistence matters. I was at Yahoo at an interesting time, around incredibly talented people, some of whom went on to start companies — and what set apart the ones that worked, in general, was persistence: the determination to make the thing work come hell or high water and find a path to get there. That’s really underrated. You could be the smartest, most technical person in the world, but without persistence, at some point it gets really hard and most people give up. And on picking problems — all startups are really hard; going from zero to one is hard no matter what you’re doing. So pick something where, if it has the impact you hope, that impact will be really large. Choosing an “easier” thing doesn’t actually make it less hard.
Daniel: If I just build it and I’m persistent, will they come?
Ariel: No, they don’t show up — I wish it were that way. You have to go after the customers. A startup usually has a product that’s 5x or 10x better than the incumbent in some vector; the real race is getting distribution faster than the incumbent gets a good product. People underestimate how hard distribution is — and that takes persistence too: persistently talking about what you’re doing, because you can tell someone once and they won’t care.
