AI Glossary / Graphs and symbolic AI / Intermediate

Perspective projection

Definition

Perspective projection maps a three-dimensional point to a two-dimensional image through a camera center. In the ideal pinhole model, the image position depends on the point's horizontal and vertical coordinates divided by its depth.

What it means

Objects farther from a camera generally appear smaller because the same physical displacement spans a smaller angle. Perspective projection describes that geometry mathematically. It is used in computer vision, graphics, robotics and camera-based measurement.

The mapping loses information: many three-dimensional points along the same viewing ray produce the same ideal image location. A single pixel location therefore does not determine a unique depth without additional information.

Perspective projection differs from orthographic projection, which uses parallel projection rays and does not shrink an object's image solely because it moves farther away. The choice of model changes which geometric assumptions are appropriate.

How it works

First express a world point in the camera's coordinate frame using the camera pose. For an ideal forward-facing pinhole convention, let the coordinates be X, Y and Z, where Z is positive depth.

Normalized image coordinates are x = X/Z and y = Y/Z. Camera intrinsics convert these to pixels. With zero skew, u = fx × X/Z + cx and v = fy × Y/Z + cy, where fx and fy are focal-length scale factors in pixels and (cx, cy) is the principal point.

Real lenses can introduce distortion, so practical calibration may estimate additional parameters. Coordinate conventions also matter: image axes, units and signs must agree throughout the calculation.

Projection identifies a geometric image location, not guaranteed visibility. A point can be outside the image boundaries, behind the camera or hidden by another surface. Occlusion and valid camera depth must be handled separately.

Three practical examples

1. Augmented reality: A system projects a virtual object's three-dimensional points into the camera image so that the overlay aligns with a modeled real-world location.

2. Robot guidance: A robot predicts where a known object should appear given its pose and camera calibration, then compares that prediction with detected image features.

3. Visual measurement: An inspection system accounts for camera geometry when relating pixel distances to physical dimensions. A simple pixels-per-millimeter conversion is not generally valid across arbitrary depths.

Think of it this way

Think of looking through a tiny hole at a scene. Each direction through the hole corresponds to a location on the image, while points farther along that same direction remain aligned.

The analogy captures central projection but omits lens distortion, finite aperture, blur and sensor effects. The ideal geometry is useful precisely because it separates those additional complications from the core mapping.

Worked example

Assume a calibrated ideal camera with fx = fy = 800 pixels and principal point (320, 240). A point in camera coordinates is (X, Y, Z) = (1, 0.5, 4) meters.

Its projected coordinates are u = 800 × 1/4 + 320 = 520 and v = 800 × 0.5/4 + 240 = 340. The point is displaced 200 pixels horizontally and 100 pixels vertically from the principal point.

Now move the point to (1, 0.5, 8). Its projection becomes (420, 290), halving those displacements because the depth doubled while X and Y stayed fixed.

By contrast, the point (2, 1, 8) projects to (520, 340), exactly like the original point. All its coordinates doubled, so X/Z and Y/Z stayed the same. This demonstrates why image position alone cannot recover depth. The example assumes the stated axis convention and omits distortion.

Illustrative case study

In a fictional project, Alder Vision estimates package width from image pixels using one conversion factor measured at a reference distance. The estimate changes when packages are placed farther from the camera, despite their physical widths remaining unchanged.

The team restricts the measurement setup to a calibrated plane or adds the depth information required by its chosen method. It also checks camera pose and distortion before interpreting image measurements as physical dimensions.

The lesson is geometric: apparent size depends on depth under perspective projection. A consistent detector does not remove that dependence or make an uncalibrated measurement physically meaningful.

Common misconceptions and mistakes

• Every image point identifies one three-dimensional location. It usually identifies a viewing ray without a unique depth.
• Doubling depth always halves the pixel coordinates. It halves offsets from the principal point when the other camera-frame coordinates stay fixed.
• Focal length in millimeters can be inserted directly into a pixel equation. Units must be converted consistently.
• Projection proves that a point is visible. Occlusion and image boundaries still matter.
• A pinhole model automatically includes lens distortion. Distortion requires additional modeling when relevant.

Business use

Understanding projection helps teams evaluate claims about visual measurement, camera alignment and three-dimensional reconstruction. Ask which coordinate frames, calibration parameters and depth assumptions support the result.

A demonstration on objects lying in one plane may not establish performance on objects at varied distances. Likewise, an overlay that works from one camera position may drift if pose estimation or calibration changes. Tests should vary the geometry that matters to the application.

Separate image-detection error from geometric-model error. A precisely located pixel can still yield a poor physical estimate if depth is wrong. Recording calibration versions, units and coordinate conventions makes these failures easier to identify and reproduce.

People also ask

Can perspective projection recover depth by itself?

No. Additional views, known geometry, depth sensing or other assumptions are needed to resolve the ambiguity.

Why do parallel railway tracks appear to converge?

Their projected image directions can approach a vanishing point, even though the tracks remain parallel in three-dimensional space.

Is perspective projection only for photographs?

No. It is also used to render virtual scenes and predict camera observations in robotics.

Check your understanding

Why do (1, 0.5, 4) and (2, 1, 8) produce the same pixel location in the worked example?

Answer: Both points have the same ratios X/Z and Y/Z, so they lie on the same viewing ray in the ideal camera model.

Related terms

• Orthographic projection
• Pinhole camera
• camera calibration
• intrinsic parameters
• extrinsic parameters
• homogeneous coordinates
• lens distortion
• depth estimation

Learn next

Project three points that lie on one ray and verify that they share an image location. Then move one point sideways without changing its depth. Track which ratios change and identify the extra information needed to infer its physical position.

More in Graphs and symbolic AI

• Performance measure - A performance measure is the criterion used to evaluate how well an agent achieves the intended task.
• Perpetual punishment - Perpetual punishment is a repeated-game strategy feature in which a detected deviation triggers a punishment phase that continues for every remaining round.
• Persistence arc - A persistence arc is a directed connection from a variable at one time step to its copy at a later step in a dynamic Bayesian network.
• Pessimistic description - A pessimistic description is a conservative description of a high-level action in angelic hierarchical planning.
• Phoneme - A phoneme is a contrastive sound category in a language: replacing it with another phoneme can distinguish one word from another.
• Phrase search - Phrase search retrieves text in which a sequence of query terms occurs with specified positional relationships, usually adjacent and in the same order for...

Sources

OpenCV documents the pinhole camera model, camera intrinsics, pose transformations and distortion parameters: https://docs.opencv.org/4.x/d9/d0c/group__calib3d.html

The pixel calculations and Alder Vision case are original illustrations using an ideal zero-skew camera and explicitly stated units.

Previous: Personalized search Next: Perturb-and-MAP sampling