Must landmarks be placed manually?

No. They can be located manually, automatically, or through an automatic proposal followed by review.

Are landmarks always visible physical corners?

No. They can be other consistently defined locations, but their interpretation and measurement method must be explicit.

Is a semilandmark identical to a fixed landmark?

No. Semilandmarks represent locations along curves or surfaces and may require methods that handle their different correspondence assumptions.

AI Glossary / Computer vision and multimodal AI / Beginner

Landmark point

Definition

A landmark point is a consistently identifiable location on an object or shape that can be matched with corresponding locations in other images or examples. In computer vision and shape analysis, landmark coordinates provide a structured way to describe geometry, align objects, measure changes, or train a system to locate meaningful parts.

What it means

A landmark is useful because its identity persists across the examples being compared. The left corner of a printed rectangle should correspond to that same corner in another image, rather than whichever visible corner happens to be closest.

In biological shape analysis, landmarks may represent anatomically corresponding locations. In engineering, they may be corners, intersections, or deliberately defined reference points. The required meaning depends on the task and must be documented.

A landmark point is not automatically the same as an arbitrary feature detector response. A visually distinctive keypoint may be convenient for matching without having a stable semantic name. Conversely, a meaningful landmark may be difficult to see clearly in some images.

How it works

Define a landmark protocol before collecting coordinates. Specify each point's identity, how to locate it, which coordinate system is used, and what to do when it is obscured or ambiguous. Consistent ordering matters because many shape models treat the ordered coordinate list as their input.

Annotators or algorithms then locate each landmark. Store visibility and uncertainty information where appropriate rather than silently substituting a guess. Repeat annotations on a sample to estimate how consistently the protocol can be applied.

Use corresponding points for alignment, geometric measurements, or statistical analysis. Translation, rotation, and sometimes scale can be removed when the goal is comparing shape rather than camera placement or size.

Validate both coordinate accuracy and correspondence identity. A point can be very precisely located yet assigned the wrong label. Points sampled along curves may require a semilandmark method instead of being treated as individually fixed anatomical correspondences.

Three practical examples

1. Product inspection: Named corners and hole centers describe a manufactured part so its geometry can be compared with a reference design.

2. Animation: Detected facial landmarks help align an illustration or estimate visible facial geometry, subject to pose and occlusion limits.

3. Shape research: Corresponding locations on leaves or other specimens form coordinate configurations for studying variation across a collection.

Think of it this way

Think of placing numbered pins on several drawings of the same type of object. Pin number three must mean the same feature in every drawing; otherwise, differences between the pin patterns become misleading.

The analogy captures correspondence and ordering. Real landmarks are measured with uncertainty, may be hidden, and may require three-dimensional coordinates. The number of pins alone does not determine whether they describe the shape adequately.

Worked example

Suppose a triangular part has landmarks A = (0,0), B = (4,0), and C = (0,3). A second image records the corresponding points as A = (10,5), B = (14,5), and C = (10,8).

Subtracting the translation (10,5) from the second configuration makes all three coordinates match the first. The side lengths are four, three, and five units in both configurations, so this example contains a position change without a shape or size change.

Now suppose the labels B and C are accidentally exchanged in the second image. The same translation no longer aligns corresponding labels, although the unlabeled triangle still has the same geometry.

This is why landmark identity is part of the data. A shape-analysis system comparing ordered coordinates can interpret inconsistent labels as real structural variation rather than an annotation error.

Illustrative case study

Fictional case: A furniture manufacturer compares photographs of decorative panels using eight landmarks. Different annotators place one point either at a rounded corner's outermost edge or at the intersection of its extrapolated straight sides.

The resulting coordinate differences appear to indicate inconsistent production. The team writes a precise point definition, adds illustrated examples, and repeats a subset of annotations. It separates measurement variability from variation in the panels themselves.

For corners hidden by packaging, the system records missing visibility instead of automatically placing a confident point. The project improves because the meaning and reliability of each landmark become clearer, not because the team simply adds more points.

Common misconceptions and mistakes

Do not assume more landmarks always produce a better model. Redundant, inconsistent, or poorly visible points can add noise and annotation cost.

Do not confuse image coordinates with physical dimensions. Camera perspective, calibration, and object pose affect the relationship between pixels and real distances.

Do not remove scale automatically when size matters to the business question. Shape-only comparison and dimensional inspection are different tasks.

Do not treat missing or estimated landmarks as equally reliable as clearly observed points. Downstream analysis should account for their status and the procedure used to supply them.

Business use

Landmarks can make a visual system more interpretable because a reviewer can inspect named locations and understand which geometric relationships support a decision. They are useful for alignment, measurement, tracking, and compact shape representations.

Their cost includes annotation design, quality control, and maintenance across new object types or imaging conditions. Before automating, establish whether people can apply the definitions consistently. Evaluate performance by landmark and by difficult conditions rather than reporting only one average distance, which can hide failure on a critical feature.

Check your understanding

Two annotators accurately mark different definitions of the same named corner. Is their disagreement necessarily a defect in the object?

Answer: No. It may arise from an inconsistent landmark protocol. The definitions and repeatability should be checked before interpreting shape differences.

Related terms

Statistical shape analysis
Image registration
point distribution model
keypoint detection
Procrustes analysis
semilandmark
annotation uncertainty

Learn next

Define five landmarks on a simple object and ask two people to annotate the same images independently. Compare coordinate differences, revise ambiguous definitions, and then study how alignment changes the interpretation of the resulting point configurations.

Sources

Bookstein, A Statistical Method for Biological Shape Comparisons (1984), primary abstract on corresponding landmark configurations: https://pubmed.ncbi.nlm.nih.gov/6738086/ . Bookstein, Morphometric Tools for Landmark Data, publisher description: https://www.cambridge.org/core/books/morphometric-tools-for-landmark-data/A18AD895945610BAD6396E76CEB512CD . Indexed primary summaries were consulted for landmark correspondence; examples here are original.

Previous: Landmark detection Next: Landmark points