Is landmark detection the same as object detection?

No. It estimates reference-point locations, while object detection typically localizes or identifies larger object instances.

Are all predicted landmarks visible?

Not necessarily. Some systems infer locations under occlusion.

Are normalized coordinates always measured in meters?

No. Their meaning depends on the output convention and must be checked explicitly.

AI Glossary / Computer vision and multimodal AI / Advanced

Landmark detection

Definition

Landmark detection estimates the locations of predefined or otherwise task-relevant reference points in an image or scene. Examples include facial points, body joints, document corners and identifiable environmental reference locations. The meaning of each landmark comes from the task's annotation scheme or geometric model, and the output may use image coordinates or an estimated three-dimensional coordinate system.

What it means

The detector supplies points with a defined role, such as a particular joint or corner, rather than merely identifying a whole object. Those points can support alignment, pose estimation, measurement or interaction. A landmark is not always a visually sharp corner, and a model may estimate its position even when it is partly occluded. Consequently, a returned coordinate should not automatically be treated as a directly observed physical location. The coordinate convention and any visibility or confidence information are important parts of the output.

How it works

Methods may locate landmarks through feature matching, statistical shape models, heatmap prediction or direct coordinate regression. A pipeline can first find a region of interest and then estimate landmarks within it. Coordinates must be mapped back through any crop, resize or rotation used during preprocessing. Some systems track results across video frames, but detection and temporal tracking are different functions. Evaluation should match the task's point definitions and account for image scale, occlusion and the consequences of localization errors. Three-dimensional outputs also require explicit interpretation of their reference frame and units.

Three practical examples

A document-processing application detects page corners to guide a subsequent geometric rectification and checks whether all corners are actually visible.
An animation prototype estimates body landmarks from video and handles missing or uncertain points before using them to control a character.
A visual-analysis team measures landmark localization error against consistently annotated reference images rather than treating every output point as equally reliable.

Think of it this way

Imagine placing labeled pins on a map: one pin marks the entrance, another a bridge and another a specific road junction. The labels tell you what the points mean, while their positions determine whether later navigation is accurate. Landmark detection makes a similar commitment to both identity and location. The analogy also shows why swapping two labels is not a small coordinate error, and why a precise-looking pin can still be placed at an inferred rather than directly visible location.

Worked example

In an original coordinate-conversion example, a detector reports normalized position (0.25,0.60) using the explicitly defined rule x_pixel=x_norm×width and y_pixel=y_norm×height. For an 800-by-500 image, the position is (200,300). If the detector actually processed a crop, its coordinates must first be interpreted in that crop's frame and then translated into the original image. The calculation assumes the stated normalization convention; some interfaces use different pixel-center or endpoint definitions. It therefore illustrates why the output contract must be checked rather than treating normalized coordinates as universally interchangeable.

Illustrative case study

A fictional interactive-learning team uses hand landmarks to position onscreen controls. The prototype works in clear frontal views but becomes unstable when fingers overlap. The team checks visibility conditions, filters unreliable outputs and validates the interaction across representative camera positions. It also verifies that cropping and mirroring are applied consistently. This scenario illustrates how a coordinate stream becomes a usable application only after additional interpretation. It does not claim that a confidence score guarantees correctness or that temporal smoothing can recover landmarks that the input does not adequately reveal.

Common misconceptions and mistakes

Landmarks are not necessarily generic keypoints: their identities often follow a specific semantic scheme. Another mistake is mixing normalized coordinates, pixel coordinates and estimated world coordinates. A third coordinate does not automatically imply calibrated physical depth. Occluded landmarks may be inferred, and a returned point is not proof of visibility. Left-right conventions can also differ between anatomical labels and screen positions. Finally, small average localization error can hide consequential failures on particular poses, lighting conditions or groups of inputs.

Business use

Landmark outputs can support alignment, interaction and visual analysis, but teams should document the point definitions and coordinate system. Validate the full path from capture and cropping through final use. Include difficult views and missing-point handling in testing. If measurements carry physical meaning, establish calibration and uncertainty rather than converting model coordinates into real-world claims without support. Avoid using a landmark detector's availability as evidence that it is suitable for every application involving the same object category. The required accuracy depends on what the application does with the points.

Check your understanding

What pixel position followed from the stipulated normalization example?

Answer: (200,300).

Why would a cropped input require another conversion?

Answer: Its coordinate origin and dimensions differ from the full image.

Why is a three-dimensional output not automatically a physical measurement?

Answer: The reference frame, scale and calibration may not support that interpretation.

Related terms

Image registration
Keypoint detection
pose estimation
coordinate normalization
visibility
feature tracking

Learn next

Define a small set of landmarks and annotate them consistently across several views. Convert coordinates between a crop and its full image, checking known points. Then examine occluded cases and separate direct visibility from model inference before using the results in an application.

Model Inference

Sources

Google, MediaPipe Pose Landmark Detection Guide: https://developers.google.com/edge/mediapipe/solutions/vision/pose_landmarker Google, MediaPipe NormalizedLandmark reference: https://ai.google.dev/edge/api/mediapipe/python/mp/tasks/components/containers/NormalizedLandmark The coordinate calculation, map-pin analogy and interactive-learning case are original educational illustrations. The entry describes the broader task and does not transfer one product's coordinate conventions to every landmark detector.

Previous: Landing zone (software) Next: Landmark point