cd ~/
IndoorLocalizationVPSChallengesMeta

Indoor Visual Positioning: Different Beast Entirely

Why indoor localization is much harder than outdoor - GPS absence, feature similarity, and the path forward.

Evyatar Bluzer
3 min read

GPS does not work indoors, and buildings look more alike than streets do. Between them, those two facts make indoor VPS a much harder problem than outdoor, and outdoor was never easy.

Why is indoor harder?

Start with the fallback problem. Outdoors, if VPS fails you fall back to 3m GPS. Indoors, if VPS fails you have nothing, so the system has to work every single time.

Then there is feature ambiguity. Outdoor features are buildings, street signs, unique architecture. Indoor features are white walls, drop ceilings, and identical corridors. Office floors are nearly identical to the floors above and below them, retail stores follow templates, and airports repeat the same gate layout over and over.

Viewpoint diversity is worse too. Outdoors you get many viewpoints from streets and tourist photos; indoors you get fewer photos from more restricted viewpoints, which means less training data and harder coverage. Lighting cooperates less as well: daylight is predictable because the sun's position is known, while artificial lighting varies by venue, flickers, and throws harsh shadows.

And scale compounds all of it. One building holds thousands of features. All buildings hold billions, many of them near-duplicates, so false matches across buildings are common.

Technical Approaches

Hierarchical localization

Narrow the search in stages, the coarse-to-fine retrieval the outdoor pipeline already uses, with the stages redrawn for buildings. First, which building - coarse, via WiFi, BLE, or an image global descriptor. Then, which floor - medium, via image retrieval within the building. Finally, where on the floor - fine, via feature matching. Each level cuts the search space before the expensive step runs.

Coarse-to-fine indoor searchFour horizontal bars stacked top to bottom, each narrower than the one above, connected by downward arrows with labels; the bottom, narrowest bar is highlighted and carries the accuracy figures beside it. All mapped buildings billions of features, many of them near-duplicates across buildings narrows to one building Which building (coarse) WiFi, BLE, or an image global descriptor narrows to one floor; multi-floor confusion is still an open gap Which floor (medium) image retrieval within the building only now does the expensive step run Where on the floor (fine) feature matching 30 cm accuracy, 85% success in mapped Meta offices today
Coarse-to-fine indoor search with schematic bar widths: building, then floor, then position, each stage cutting the candidates before the expensive feature-matching step runs.

Semantic features

Generic features drown in all that repetition, so we add semantic understanding. Knowing "this is a Starbucks" narrows the candidate locations dramatically, room types like bathroom, elevator, or conference room provide context, and signage and text are highly distinctive. The plan combines geometric and semantic features.

Multi-modal fusion

Visual alone is not enough indoors. We fuse it with WiFi fingerprinting for room-level accuracy, BLE beacons where they exist, magnetic field signatures, and pedestrian dead reckoning. Each modality contributes something, and fusion covers for whichever one is failing at the moment.

Mapping Indoor Spaces

The mapping problem is as much social as technical. Access is controlled, so nobody gets to walk into an office and start capturing, privacy sensitivity is high (people, screens, documents), the environment is dynamic because furniture moves daily, and there is more indoor space than outdoor surface to cover. Our approaches: partner with venue operators, offer mapping-as-a-service for businesses, use robotics for systematic capture, and take crowd-sourced imagery only with strong privacy controls (the outdoor defenses are the floor, not the ceiling).

Current State

Indoor VPS works today in controlled environments, meaning Meta offices, with 30cm accuracy in mapped areas and an 85% success rate. That is not ready for external launch. In June I put indoor support on the three-month list. The three months are up this week, and indoor works in Meta offices. The remaining gaps are multi-floor confusion, similar-room discrimination, and mapping efficiency. We are targeting an indoor beta in 2022.

Comments