AI Glossary / AI and machine learning / Intermediate
Landmarks
Definition
Landmarks are distinctive, locatable points used as reference markers: in vision, named positions on faces or objects; in search and planning, well-known nodes that help estimate distances.
What it means
Two communities use the word for the same idea. In computer vision, landmarks are the named points on a structure: eye corners, nose tip, lip edges on a face, or the corners and features of a manufactured part.
In route planning and graph search, landmarks are preselected well-known nodes: by storing distances to a handful of landmarks, the system can estimate any other distance quickly, the way you might find your way using knowing how far places are from the city center.
Both uses convert a hard spatial problem into simpler references. A face becomes sixty-eight coordinates; a road network becomes a table of distances to a few hubs. Everything downstream, alignment, measurement, estimation, builds on the references.
Quality depends on consistency: a landmark that means slightly different pixels to different annotators, or a routing landmark placed in a cul-de-sac, degrades every computation that trusts it.
For a business, landmarks appear inside face analytics, document processing, mapping and logistics tools. The buyer's question is which landmarks, defined how precisely, and measured how accurately on your data.
In some pipelines the landmarks are the product: once the coordinates exist, face alignment, expression analysis, gaze estimation and document deskewing are all downstream arithmetic on the same pins.
How it works
Define the landmark set precisely: named points with written placement rules that two people would mark identically.
Label training data with the landmarks and validate detector accuracy on imagery matching the deployment.
In routing, choose landmark nodes that are well-connected and spread out, and precompute distances to them.
Use the references downstream: alignment and measurement in vision, distance estimation in search.
Audit drift: re-validate landmark accuracy when imagery, cameras or the road network change.
Three practical examples
- A face-analysis system placing sixty-eight facial landmarks to align portraits before measuring anything.
- A logistics platform estimating delivery times from precomputed distances to a dozen hub landmarks.
- A factory vision system locating screw-hole landmarks on a component to check assembly alignment.
Think of it this way
Think of giving directions by pub and church spire rather than by coordinates: meet me two streets past the spire, opposite the old pub. Landmarks are the spires, human or algorithmic: fixed references that make every other location describable.
The limit: directions fail when the spire is demolished or the pub becomes a flats. A landmark set that stops matching reality, a changed road layout, a restyled product, silently corrupts everything referenced to it.
Worked example
aligning a million portraits
Suppose a bank verifies identity by comparing a selfie with an ID photo. The figures are illustrative.
Facial landmarks on both images align them to a standard pose before comparison, removing angle and distance as confounders.
Match accuracy rises nine points over unaligned comparison, and false rejections, the customer-facing cost, fall by half.
The landmarks never decide anything themselves; they simply put both faces on the same map, which is what makes the comparison fair.
Illustrative case study
A fictional routing startup placed its routing landmarks at the most photographed locations, reasoning that famous equals useful, and found its distance estimates were wildly optimistic in the suburbs.
The landmarks clustered in the center, so suburban routes, far from every reference, got estimates no better than guesses.
The invented lesson: a landmark earns its place by coverage, not fame. Spread the references across the territory they must describe.
The platform relaunched with landmarks chosen by graph coverage, ten well-spread hubs, and suburban estimates tightened immediately, without any change to the routing algorithm itself.
Common misconceptions and mistakes
- Landmarks and keypoints are different things. They are the same idea under different traditions: vision says landmarks or keypoints, graphics and planning say landmarks.
- More landmarks always help. Each is a labeling and maintenance cost; the right set covers the structure with the fewest well-chosen points.
- Landmarks are permanent once chosen. Imagery, products and road networks change, and the reference set needs periodic re-validation like any other asset.
Business use
Specify landmark definitions in writing before any labeling starts; vague points produce inconsistent data.
Re-validate landmark accuracy on current imagery before any downstream metric is trusted.
In mapping work, choose references for territorial coverage and connectivity, not for prominence.
When buying face or document analytics, ask for the landmark error distribution, not just the mean: the tail of bad detections is where the downstream failures live.
People also ask
What is a facial landmark?
A named point on the face, such as an eye corner or nose tip, detected as coordinates and used for alignment and measurement.
How do landmarks speed up route planning?
Precomputed distances to a few well-placed nodes let the system bound any other distance quickly, pruning the search.
How many landmarks does a task need?
The fewest that pin down the structure: dozens for a face, four for a document corner set, a spread handful for routing.
Who decides where a landmark sits?
The annotation guide: a written definition precise enough that two trained people mark the same pixel, which is why the guide is the real asset.
Check your understanding
Why do landmarks make downstream measurements more reliable?
Answer: Because they provide fixed, named references, so alignment and measurement are computed against stable points rather than against raw, unregistered images or locations.
Related terms
Learn next
AI for Business Owners covers scoping a vision or mapping vendor's claims to the specific references you are buying: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Labeled example - A labeled example is a single training case with its known answer attached: the inputs plus the label, the atomic unit of supervised learning.
- Lagged features - Lagged features are inputs built from the past values of a series: yesterday's sales, last week's balance, the same customer a month ago, giving a model...
- Lambda - Lambda is the traditional symbol for the regularization strength dial: the number that decides how heavily a model is penalized for large weights, trading...
- Large margin nearest neighbor - Large margin nearest neighbor, or LMNN, is a supervised metric-learning method that learns a distance transformation for nearest-neighbor classification.
- LARS Lasso - LARS Lasso is the lasso fitted by the Least Angle Regression algorithm: an efficient method that traces the whole path of models as the penalty relaxes...
- Lasso - Lasso is regression with an L1 penalty, famous for driving the weights of unneeded fields to exactly zero, so fitting the model and selecting its fields...
Sources
[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Landmark points Next: Language