ComparisonAugmented Reality (AR)

Marker, image, model target or markerless: choosing AR tracking for equipment

AR tracking methods differ in what they need from the site and how well content stays aligned once the worker moves. Fiducial markers and image targets are simple and dependable but must be placed and maintained; CAD-based model targets need good 3D data; markerless SLAM needs nothing prepared but drifts. Most industrial apps combine two. Here is how they compare, how to choose and how to test them on site.

Reviewed 8 min read

On this page
  1. Five tracking approaches rated on site conditions
  2. Why a small misalignment changes what a technician does
  3. Tracking vocabulary used in SDK documentation
  4. What each method asks of your site and your data
  5. A field test protocol for alignment accuracy
  6. Matching a tracking method to the asset and the job
  7. Hybrid set-ups that combine two methods
  8. Questions and answers
  9. Sources

Five tracking approaches rated on site conditions

CriterionFiducial marker or QR codeImage targetModel target from CADMarkerless SLAMSpatial or cloud anchor
What it needsA printed marker fixed at a known spot on the assetA flat, feature-rich graphic already on or near the assetAn accurate 3D model or scan of a rigid objectNothing prepared; the device maps surfaces as it goesAn area mapped and saved earlier
Initial alignmentPrecise if the marker position is recordedPrecise while the image is clearly in viewPrecise once the user matches the guide viewApproximate; content is placed by hand or on a detected planeDepends on how well the area was mapped
Drift as the user movesLow near the marker, growing with distanceLow near the image, growing with distanceLow while the object stays in viewAccumulates with distance and timeCorrected whenever the device relocalizes
Setup effortLow per asset, but every asset needs a markerLow if suitable graphics already existHigh at first: model preparation and trainingLowestModerate: map each location and manage anchor lifetimes
Glare, low light, shiny or plain surfacesTolerant if the marker is clean and matteStruggles with glare and worn graphicsShiny surfaces are unsupported in some tools3Struggles on featureless or reflective surfacesStruggles in the same conditions as SLAM
When the equipment changesMove or replace the markerReplace the target if the graphic changesUpdate the model and retrainNothing to updateRe-map if the surroundings change a lot

Ratings are qualitative and describe typical behavior. Accuracy depends on the device, the SDK and the environment, so measure it with the protocol further down this page.

Why a small misalignment changes what a technician does

In consumer AR, a virtual object that sits slightly off is a cosmetic flaw. On equipment, an arrow that lands on the wrong terminal, fastener or valve is an instruction to do the wrong thing. Workers notice quickly: once an overlay has misled them, many stop trusting it and return to the manual, and the project loses its value without anyone formally rejecting it.

The tolerance you need comes from the task. Showing which cabinet to open can survive a large error; pointing to one connector in a dense row cannot. Write the tolerance down per step before choosing a method, in units a technician could check with a ruler. Where no method meets it reliably, design the step so the overlay frames an area and the worker confirms the exact part. ColdAI's AR work keeps conventional verification for safety-critical steps for the same reason1.

Tracking vocabulary used in SDK documentation

Registration
Aligning the coordinate frame of digital content with a physical object, so a point in the model coincides with the same point on the equipment.
Drift
Gradual error in the estimated pose as small tracking mistakes accumulate, seen as content sliding away from where it was placed.
Relocalization
Recognizing a previously mapped place and snapping the device's estimated position back onto that map, which also removes accumulated drift.
Guide view
A silhouette of the object from a set angle and distance, which the user lines up with the real object so that model tracking can start3.
Occlusion
Real objects passing in front of digital content; good handling hides the content behind hands, tools and pipework instead of drawing over them.

What each method asks of your site and your data

Fiducial markers and QR codes are high-contrast patterns designed to be detected quickly. Their weakness is logistics: someone must print, fix and maintain a marker on every asset at a recorded offset from the equipment, and replace it when it is painted over, damaged or removed during maintenance. OpenXR now includes a cross-vendor extension for tracking markers such as QR codes, which reduces dependence on any one SDK2.

Image targets use graphics that already exist, such as a nameplate, label or control panel overlay. ARCore's guidance asks for flat images with many distinctive features, warns against repetitive patterns and asks that the image fill a good part of the camera frame for detection4. ARKit offers equivalent image detection and tracking on Apple devices5. Worn or glossy labels are the usual failure.

Model targets recognize the object by its shape, from a 3D CAD model or scan. Vuforia's documentation requires the object to be rigid with stable surface features and says shiny surfaces are not supported3. On Apple Vision Pro, a reference object is trained in Create ML from a USDZ model, and Apple advises that objects which are mostly stationary, rigid and non-symmetrical work best6. Preparation is real work: simplifying geometry, matching colors to the physical object and retraining after modifications.

Markerless SLAM maps feature points and planes as the device moves, which is why it needs no preparation and why it drifts. It suits content that only has to sit roughly on a floor or bench. Spatial and cloud anchors save a mapped place so content returns between sessions; persistence and sharing across devices are covered in shared spatial anchors.

A field test protocol for alignment accuracy

Run the same protocol for every candidate method, on the real equipment, with the devices you intend to deploy.

  1. Mark check points

    Choose several points the task depends on, such as a terminal, a bolt head and a valve handle, and record their positions from the CAD model or with a tape measure.

    Output
    Check-point list

  2. Define the conditions

    Plan sessions in normal light, low light and strong glare, on the reflective or plain surfaces the site really has, with the dirt and wear of a working plant.

    Output
    Test matrix

  3. Measure initial error

    After alignment, place a virtual pin on each check point and measure the gap to the real point with a ruler, from the distance a technician would actually work at.

    Output
    Initial error per point and condition

  4. Measure drift

    Walk around the asset, step back, crouch and return. Measure each check point again; the change from the first reading is drift.

  5. Test occlusion and interruption

    Put hands, tools and a second person in view, cover the marker or target, and note whether content holds, jumps or disappears.

  6. Repeat and record the spread

    Repeat the full run on different days with different people. Consistency matters as much as the best result.

    Output
    Accuracy and repeatability summary
    Owner
    AR technical lead

  7. Time the setup

    Record how long it takes from opening the app to usable tracking, failed attempts included. A precise method that takes too long to start will be skipped.

Matching a tracking method to the asset and the job

  • If

    You have many identical assets and can fix a label to each one.

    Then

    Use fiducial markers or QR codes, ideally encoding the asset ID so the right content loads automatically.

    Detection is fast and dependable, and the marker doubles as asset identification.

  • If

    The asset has accurate CAD data, is rigid and has a matte finish.

    Then

    Use a model target, so there is no marker to fix or maintain.

    Shape recognition survives label damage and works from the angles the task needs, once the model is prepared.

  • If

    The asset is shiny, transparent, flexible or modified often.

    Then

    Avoid model targets; put a marker on a stable part of the asset or on a fixed fixture nearby.

    These are exactly the conditions in which shape recognition fails.

  • If

    Content only needs to sit approximately in a room, such as a keep-out zone on the floor.

    Then

    Use markerless plane detection, adding anchors if the content must return later.

    The tolerance is loose enough that drift will not mislead anyone.

  • If

    Content must reappear in the same place for different people and devices.

    Then

    Use persistent or shared anchors, with a marker as the fallback for initialization.

    Anchors provide persistence, and the marker gives a known starting point when relocalization fails.

Hybrid set-ups that combine two methods

Questions and answers

Do we need CAD models to align AR content to our equipment?

Only for model targets. Markers, image targets and markerless tracking need no CAD data, and many work instruction apps run on markers alone. If you want shape-based recognition, existing CAD is the usual starting point, but it often needs simplifying and color-matching first, and some tools accept a 3D scan where no CAD model exists.

How accurate can AR alignment on equipment be?

It depends on the method, the device, the working distance and the conditions, so treat any single figure with caution. Methods that recognize the object directly are usually more precise near it than markerless tracking, and accuracy falls as the worker moves away from whatever was recognized. Measure it with your own check points under site conditions, and design steps that need more precision so the worker confirms the part.

What happens to AR tracking when the equipment is modified?

Markers survive modifications if their mounting point is unchanged, though their recorded offset may need updating. Image targets break if the graphic is replaced. Model targets need the model updated and retrained when the shape changes noticeably. Link each asset's tracking data to its asset record so that a modification triggers a review of the AR content.

Does AR tracking work on shiny metal or stainless steel?

Poorly, for most methods. Reflections confuse feature-based SLAM and shape recognition alike, and some model target tools state that shiny surfaces are not supported. Practical workarounds are a matte marker on a stable part, tracking a nearby non-reflective fixture, or anchoring content to the room rather than to the object itself.

Sources

  1. Augmented Reality (AR): ColdAI's approach to object-aligned AR — ColdAI
  2. OpenXR Spatial Entities Extensions Released for Developer Feedback — The Khronos Group · checked 10 October 2026
  3. Model Targets — PTC Vuforia Engine developer library · checked 10 October 2026
  4. Augmented Images overview — Google ARCore · checked 10 October 2026
  5. Tracking and altering images — Apple Developer · checked 10 October 2026
  6. Explore object tracking for visionOS (WWDC24) — Apple Developer · checked 10 October 2026

More in Augmented Reality (AR)

Back to Augmented Reality (AR)

Next step

Test tracking on your own equipment before you commit

Send photos of the assets, the CAD data you hold and a description of the site conditions, and we can scope a tracking test that measures alignment and drift on the devices you plan to use.

Scope a tracking test