Plain definitions for the vocabulary that comes up when a camera, a robot or a headset needs to know where it is. Each entry says what the term means and, where it matters, when the thing is actually the right choice.
Written for people evaluating positioning technology rather than for people who already build it. Where two terms are routinely confused, the difference is stated explicitly rather than implied.
Positioning fundamentals
Visual positioning system (VPS)
A system that works out where a camera is by comparing what it sees against a prebuilt 3D map of the space. It returns full position and orientation in that map's coordinate frame. Unlike satellite positioning it works indoors, and unlike beacons it needs nothing installed in the building. See how the VPS works.
6-DoF pose
Six degrees of freedom: three numbers for position and three for orientation. A pose with only position tells you where something is but not which way it faces, which is enough to draw an arrow on the floor and not enough to put a label on the correct object.
Localization
The act of determining a device's pose within a known map. Distinct from mapping, which builds the map in the first place, and from navigation, which plans a route once the pose is known.
Relocalization
Recovering pose after it has been lost or after a session restarts. The measure that matters in practice is how long it takes and how reliably it succeeds from a viewpoint nobody composed.
Absolute versus relative positioning
The distinction that decides most deployments. Relative positioning reports movement from wherever a device started, so two devices never agree and nothing survives closing the app. Absolute positioning reports a pose in a shared frame that already exists, so content stays put and multiple devices and robots agree with each other.
Coordinate frame
The reference system a position is expressed in. Two correct positions in different frames are not comparable, which is why a robot fleet, a building model and a technician's phone can each be right and still disagree about where something is.
Drift
Accumulated error in a position estimate that grows over time or distance travelled. Any system that integrates motion without observing the world drifts without bound. An absolute fix resets the error rather than slowing its growth.
First lock
The time from opening an application to a trusted position. Often the difference between a pilot that gets adopted and one that quietly stops being used, because people under time pressure abandon anything that makes them wait.
Confidence value
A score returned alongside a pose indicating how much to trust it. Any serious system returns one, and a correct integration gates on it, because a wrong position stated confidently is worse than no position at all.
Persistent anchor
A piece of content fixed to a real-world location that survives sessions, devices and time. Requires absolute positioning against a shared map. Anchors placed against relative tracking do not come back.
Shared coordinate frame
One frame that people, robots and authored content all resolve into. Turns every handoff between them from an integration project into a lookup.
How positioning is done
GPS and GNSS
Satellite positioning. Typically three to ten metres outdoors with a clear sky, and unusable indoors or under structure. The failure mode is rarely a clean absence of signal; a receiver under partial sky reports a position that exists and is wrong.
GNSS-denied environment
Any space where satellite positioning is unavailable or untrustworthy: indoors, under decks and pipe racks, inside vessels, urban canyons where signals arrive by reflection, and areas subject to jamming or spoofing. Indoors this is the normal condition, not an edge case.
SLAM
Simultaneous localization and mapping. Builds a map as it moves and tracks against it. Excellent for a device navigating on its own; the coordinate frame starts wherever the device started, so it is private to that session and drifts. Complementary to a VPS rather than competing: SLAM handles frame-to-frame motion, the VPS supplies the absolute fix.
Visual SLAM
SLAM using camera images as the primary sensor. Cheap in hardware, dependent on visible texture, and degrades in featureless or poorly lit space.
LiDAR SLAM
SLAM using laser range measurement. More geometrically precise than visual SLAM and less dependent on lighting, at the cost of a LiDAR sensor on every device.
Visual-inertial odometry (VIO)
Combining camera motion with inertial sensors to estimate movement. What the tracking in a modern phone or headset is built on. Smooth and fast, and relative, so it drifts.
Dead reckoning
Estimating current position by integrating motion from a last known fix. Position is the double integral of acceleration, so a small constant sensor bias becomes an error growing with the square of elapsed time. Fine for seconds, not for minutes.
Ultra-wideband (UWB)
Radio positioning using precisely timed pulses between fixed anchors and tags. Ten to thirty centimetres is realistic and genuinely good. The cost is that it is an infrastructure project, and untagged things stay invisible.
BLE beacons and Wi-Fi positioning
Position inferred from radio signal strength. One to five metres, degrading as a space fills with people and inventory. Cheap per unit, with the real cost in installation and battery replacement across hundreds of devices.
RTLS
Real-time location system. An umbrella term for anything tracking the live position of people or assets, whether by radio, camera or tag. Naming the underlying method matters more than the acronym.
Fiducial marker and QR positioning
Printed targets at surveyed positions. Exact at the marker and drifting between them. The maintenance burden scales with the number of things you want to identify, which is the wrong way round.
Sensor fusion
Combining several positioning sources so each covers the others' weaknesses. The standard production arrangement is on-device tracking for smooth motion plus periodic absolute fixes to correct the accumulated drift.
Capture and formats
Reality capture
The general practice of recording a physical space as 3D data, whatever the instrument. Most industrial buildings have been captured at least once already, usually for engineering or facilities rather than for positioning.
Point cloud
A set of measured 3D points, each with a position and often colour. The rawest useful form of a scan. Accurate and heavy, and not directly viewable as a surface.
E57
The open interchange format for point cloud data, exported by effectively every professional scanner. Because it is vendor-neutral, it is usually the file a building already has. See E57 to VPS.
MatterPak
Matterport's downloadable bundle of a captured space, including a point cloud and mesh. One of two Matterport exports that work as a localization input, the other being E57.
Mesh
A surface built from triangles, usually with a texture applied. What most people picture as a 3D model. Lighter than a point cloud and easier to view, with measurement fidelity depending on how it was derived.
GLB and PLY
Common 3D file formats. GLB carries a textured mesh in a single file. PLY carries either a mesh or a point set, and is also the usual container for a Gaussian splat.
Photogrammetry
Reconstructing 3D geometry from overlapping photographs. Needs no special sensor and generally produces metric scale when the capture is set up for it.
Gaussian splatting (3DGS)
A reconstruction method that represents a scene as many small oriented blobs rather than as surfaces. Photoreal and fast to render. Not metric-scaled by default, which is the single property that decides whether it can be used for localization. See 3DGS to VPS.
NeRF
Neural radiance field. An earlier neural reconstruction approach, strong on view synthesis and generally too slow to render in real time. Largely superseded by Gaussian splatting for practical work.
Metric scale
The property that one unit in a reconstruction equals one real-world metre. A scale-free reconstruction can look perfect and still be unusable for positioning, because pose error grows with distance. A map that is not metric is a picture of a place.
Equirectangular
The projection a 360 camera writes, where a full sphere is flattened into a single rectangular frame. The format matters more than the camera badge: anything writing standard equirectangular MP4 can be ingested.
LiDAR
Distance measurement by timing reflected laser light. Found in survey scanners and in recent phones and tablets. Provides true scale directly rather than inferring it.
Coverage overlap
How much consecutive viewpoints share. The most common cause of a map that localizes in some places and fails in others, and invisible in a rendered preview. Matters more than scanner grade.
Registration
Tying individual scans into one coordinate system. Unregistered scans are separate coordinate systems, so a folder of stations is not yet a map.
MultiSet platform terms
Scan-agnostic
Accepting capture data from any source rather than requiring a proprietary device or workflow. In practice this is what removes the capture cost from a deployment, because the building has usually been scanned already.
MapSet
Multiple maps stitched into one continuous coordinate system, across floors, buildings and indoor-to-outdoor transitions. Lets a device move between spaces without re-localizing or hitting map boundaries.
Map versioning
Re-scanning a space over time while authored content stays anchored. Each new capture aligns to the active version, and anchors carry forward rather than being placed again by hand. See map versioning.
Partial map update
Recapturing only the zones that changed rather than the whole site. The question to ask any vendor, because a system needing a full recapture after every layout change will fall out of use inside a year.
Object tracking
Recognising and tracking a specific physical object from a 3D model of it, rather than positioning within a room. Answers where a part is relative to the camera right now. See object tracking.
Markerless
Working from the appearance and geometry of the thing itself, with no printed target attached. Removes the placement and maintenance burden that makes marker systems fail at scale.
Gen2
The current generation of MultiSet's localization pipeline, with improved accuracy and fewer confident-but-wrong results than the previous generation.
Deep Search
A localization mode that trades a little latency for a much better chance of succeeding from a difficult first view, such as weak texture, motion blur or partial overlap.
GeoHint and HintPosition
Supplying an approximate starting position, typically from GPS or a previous session, so the system searches a smaller region of the map and locks faster.
360 Virtual Tour
A navigable panorama walkthrough generated from the same capture that produced the localization map. The human-readable output alongside the machine-readable one.
Active and archived maps
An active map is live and available to localize against. An archived map is retained for rollback or later reactivation but is not served. The distinction matters because plan limits usually count them separately.
One capture, two artifacts
The idea that a single pass through a building should produce both a machine-readable map devices localize against and a human-readable reconstruction people look at. Most tools produce one or the other, and the second site visit is the expensive part.
Deployment and sovereignty
Private cloud and VPC
Running in a single-tenant environment inside your own cloud account. Data stays within infrastructure you control, which usually satisfies procurement policy and data residency requirements.
Self-hosted and on-premises
Running on your own hardware inside your own network. The usual answer where a third-party cloud call is not permitted at all.
On-device
Localization computed on the device itself with no network call. Works offline, keeps imagery local, and gives the lowest latency.
Air-gapped
No network path to the outside at all, as a physical property rather than a firewall rule. Requires software that runs with no outbound call including licence checks, updates delivered as signed offline artifacts, and pinnable versions.
Data sovereignty
The principle that data is subject to the laws of the country it is physically stored in. A legal condition rather than a technical one, and the reason interior scans of sensitive facilities are treated carefully. See deployment modes.
Data localization
The requirement that data be stored inside a specific country's borders. Not the same as sovereignty: a hyperscaler region inside your country satisfies localization while the operator may still be subject to foreign disclosure law.
Applied and industry terms
Indoor navigation
Guiding a person or machine inside a building where satellite positioning does not reach. Splits into two different problems: routing someone to a room, where metres are enough, and identifying one object among identical ones, where centimetres are the requirement.
Indoor positioning
Determining position inside a building, by any method. Positioning is the measurement; navigation is what you build on top of it.
Spatial computing
Software that understands and acts on physical space rather than on a screen. Positioning is its precondition: nothing else in the stack works if the device does not know where it is.
Physical AI
AI systems that perceive and act in the physical world rather than only processing data about it. Perception and planning have improved faster than the positioning layer beneath them, which is why an absolute, shared pose is increasingly the bottleneck. See robotics.
Spatial intelligence
A machine's ability to reason about where things are and how a space is arranged. Depends on a persistent map rather than on what a single sensor can see right now.
Digital twin
A maintained digital representation of a physical asset or facility. Usually built for viewing, measuring and documenting. A digital twin does not by itself tell a device where it is; that requires a localization map derived from the same reality.
BIM
Building information modelling. The designed model of a building and its systems. Comparing BIM against as-built reality is one of the clearest uses of accurate positioning, because the overlay has to land on the actual component.
As-built
The record of how a facility was actually constructed, as opposed to how it was designed. Usually captured by laser scanning, and usually already sitting on a drive.
Georeferencing
Tying a local map to real-world coordinates. Where a site has surveyed control points, they fix the map to the world and the VPS fixes the device to the map. The two are complementary, not alternatives.
RTK
Real-time kinematic positioning. A correction technique giving centimetre-level satellite accuracy outdoors with a base station. Useful for georeferencing a map, and still unavailable indoors.
AMR and AGV
Autonomous mobile robots and automated guided vehicles. Both already localize competently, in their own private coordinate frame, which is why they and the people around them often disagree about where things are.
ROS 2
The standard robotics middleware. Positions are shared through its transform tree, so an absolute pose published there arrives alongside odometry rather than as a parallel system to reconcile.
Smart glasses
Head-worn displays with cameras. They change the positioning requirement rather than just the form factor: nobody frames the shot, queries become continuous, and a few degrees of heading error is visible in a way it never was on a phone. See SDKs.
WebXR
A browser standard for AR and VR. Lets an experience run from a link or QR code with no app install, which matters most for pilots and for anything customer-facing.
Planogram compliance
Checking that the physical shelf matches the published plan: right SKU, right bay, right facing count. A positioning problem in disguise, because a photo is only useful if you know exactly which bay it shows.
AR work instructions
Procedural guidance overlaid on the equipment it describes. Pays back where the overlay must land on a specific unit among units that look the same. Where a general area is enough, a printed sheet is cheaper.
Missing a term, or think a definition here is wrong? Tell us and we will fix it.
