cd ~/
IndoorLocalizationVPSChallengesMeta

Indoor Visual Positioning: Different Beast Entirely

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

Evyatar Bluzer
2 min read

Outdoor VPS is hard. Indoor VPS is harder. GPS doesn't work indoors, and buildings look more alike than streets do.

Why Indoor is Harder

No GPS Fallback

Outdoors: If VPS fails, fall back to 3m GPS. Indoors: If VPS fails, you have nothing.

The system must work, always.

Feature Ambiguity

Outdoor features: Buildings, street signs, unique architecture. Indoor features: White walls, drop ceilings, identical corridors.

Office floors are nearly identical to other floors. Retail stores follow templates. Airports have repetitive gates.

Limited Viewpoint Diversity

Outdoor capture: Many viewpoints from streets, photos from tourists. Indoor capture: Fewer photos, more restricted viewpoints.

Less training data, harder coverage.

Lighting Variability

Outdoor: Daylight is predictable (sun position). Indoor: Artificial lighting varies by venue, flickers, creates harsh shadows.

Scale Challenges

One building: Thousands of features. All buildings: Billions of features, many duplicates.

False matches across buildings are common.

Technical Approaches

Hierarchical Localization

First: Which building? (Coarse - WiFi, BLE, or image global descriptor) Then: Which floor? (Medium - image retrieval within building) Finally: Where on floor? (Fine - feature matching)

Each level reduces search space.

Semantic Features

Instead of generic features, use semantic understanding:

  • "This is a Starbucks" narrows locations dramatically
  • Room types (bathroom, elevator, conference) provide context
  • Signage and text are highly distinctive

Combining geometric and semantic features.

Multi-Modal Fusion

Visual alone isn't enough. Fuse with:

  • WiFi fingerprinting (room-level accuracy)
  • BLE beacons (if present)
  • Magnetic field signatures
  • Pedestrian dead reckoning

Each modality contributes. Fusion handles individual failures.

Mapping Indoor Spaces

Challenges:

  • Access control (can't just walk into offices)
  • Privacy sensitivity (people, screens, documents)
  • Dynamic environments (furniture moves daily)
  • Scale (more indoor space than outdoor surface)

Approaches:

  • Partner with venue operators
  • Mapping-as-a-service for businesses
  • Robotics for systematic capture
  • Crowd-sourced with strong privacy controls

Current State

Indoor VPS in development:

  • Working in controlled environments (Meta offices)
  • Accuracy: 30cm in mapped areas
  • Robustness: 85% success rate

Not ready for external launch. Key gaps:

  • Multi-floor confusion
  • Similar-room discrimination
  • Mapping efficiency

Targeting indoor beta in 2022.

Comments