Crowdsourced Mapping: Scaling Through Users
Enabling users to contribute to mapping - the technical architecture, privacy model, and quality challenges.
Professional mapping, even the partner-network version, does not scale. There are millions of places people want AR experiences, and we cannot map them all ourselves - no team could. The only path to coverage at that scale runs through the users.
The Vision
Any Quest user can contribute to mapping. Opt in to the contribution program and the device captures during normal use. Captures get uploaded with privacy preserved, aggregated into maps, and those maps become available to VPS. Everyone gets the benefit of the collective map.
Privacy Architecture
If users hand us data, they need strong guarantees in return, and the architecture has to enforce those guarantees rather than promise them.
On-Device Processing
Raw images never leave the device. Features are extracted on-device, identifying information like faces and screens is removed, and what remains is compressed to a minimal representation before upload.
Differential Privacy
Nobody should be able to determine whether a specific user contributed, the same guarantee the query side already makes. Contributions are aggregated before storage, noise is added to prevent individual identification, and minimum contribution thresholds apply.
User Control
Contribution is a clear opt-in, never a default. Users can see what they have shared, delete their contributions, and set geographic limits so the area around home never gets mapped.
Quality Challenges
User contributions vary wildly: motion blur from casual capture, poor lighting, obstructed views, duplicate coverage. The quality pipeline runs four stages - filtering rejects unusable captures, scoring ranks the rest, selection picks the best for mapping, and validation verifies accuracy before anything gets published. 90% of contributions get filtered out; the 10% that survive are gold.
Why would anyone contribute?
Not for money - we deliberately avoid payment because it attracts people gaming the system - and never because they have to, which would be an ethics problem. What is left: intrinsic motivation to help build the AR future, social recognition for contributions, plain utility in the form of better maps where you care about them, and some gamification through achievements and progress bars. The 2020 long-tail plan leaned on capture incentives for under-mapped areas. Two years on, the incentive we ruled out is the obvious one.
Coverage Strategy
Crowdsourced maps complement professional ones. Professional crews cover high-value, high-traffic locations; crowdsourcing covers the long tail of user-specific places. User contributions fill the gaps professionals cannot reach economically.
Results So Far
The beta program has 10,000 users. They have sent 1M+ contributions, of which 50K passed the quality filter, producing 100 newly mapped locations we could not have covered otherwise. 78% of participants would recommend the program. We are scaling to general availability in 2023.
Learnings
Privacy concerns are real - many users opt out over privacy worry, and trust has to be earned before scale is possible. The quality bar matters more than volume; low-quality maps hurt more than they help. Feedback loops work: users who see their contributions actually used contribute more. And the long tail is long. Most contributions come from a small set of users, and keeping everyone else engaged is the part we have not solved.