Is scaled orthographic projection the same as ordinary orthographic projection?

It adds an overall scale linked to representative viewing depth; terminology sometimes groups both under affine camera models.

Can the scale change between images?

Yes. The representative distance, and therefore the scale, can differ between views.

Does the approximation recover each point's depth?

No. Replacing individual depths by a common value does not supply those missing depths.

AI Glossary / Graphs and symbolic AI / Intermediate

Scaled orthographic projection

Definition

Scaled orthographic projection is a camera approximation that projects three-dimensional points orthographically and applies a shared scale based on a representative depth. Also called weak perspective, it replaces each point's individual perspective depth with a common value.

What it means

In a perspective camera, a point's image location depends on its distance from the camera. Two equally wide objects can appear different sizes when one is nearer. Scaled orthographic projection simplifies this relationship for a compact object whose depth variation is small compared with its distance from the camera.

The model still allows the object's overall image size to change with its representative distance. What it removes is the separate depth-dependent scale of each point within that object.

This approximation appears in computer vision because it can simplify geometric fitting and reconstruction. It is a modeling choice, not a claim that the physical camera has stopped producing perspective effects. Close views or large depth variations may expose the difference.

How it works

Express points in camera coordinates, after accounting for the object's orientation and position. In a simple camera with equal focal scaling in both image directions, perspective gives image offsets u=fX/Z and v=fY/Z relative to the principal point.

Replace each Z by a representative depth Z0. The equations become u=sX and v=sY, where the shared scale s=f/Z0. Image translation can then place the result at its appropriate location in the image.

A three-dimensional rotation still matters: turning an object changes its camera-coordinate X and Y values, even though the final projection discards individual depth scaling. Scaled orthographic projection is therefore not the same as simply ignoring orientation.

Assess the approximation against the required image accuracy. Small relative depth variation supports it, while field of view and placement also affect practical error. Lens distortion and calibration errors are separate issues; replacing depth by an average does not correct them.

Three practical examples

A vision researcher fits a distant object's approximate pose using a simpler shared-scale camera model before testing a fuller perspective model.
An animation workflow estimates a compact subject's image placement when its depth variation is small relative to viewing distance.
An inspection team compares predicted feature locations from weak perspective and perspective to decide whether the approximation meets its pixel-error tolerance.

Think of it this way

Think of photographing a shallow display from far enough away that its front and back are nearly the same distance from the camera. Treating the display as having one representative depth may preserve its image layout reasonably well.

The analogy has a limit: “nearly” must be judged against the task's tolerance. A shallow display photographed very close up can still show perspective differences that matter.

Worked example

Take focal length f=800 pixels and representative depth Z0=10 meters. The common scale is 800/10=80 pixels per meter. Consider two points with X=1 meter and Y=0, at depths 9.8 and 10.2 meters.

Scaled orthographic projection places both at horizontal offset 80 pixels. Full perspective places the first at 800/9.8, approximately 81.63 pixels, and the second at 800/10.2, approximately 78.43 pixels. The approximation errors are about 1.63 and 1.57 pixels.

Now move the first point to depth 5 meters while retaining the same representative depth. Its perspective offset becomes 160 pixels, but the approximation still predicts 80. That large discrepancy shows why a single scale can fail when depth variation is substantial.

These calculations omit principal-point translation because adding the same translation to both predictions does not change their difference. Whether the smaller errors are acceptable depends on the application, not on the model's name.

Illustrative case study

In a fictional project, Maple Vision measures feature positions on a small component viewed from a fixed camera. A shared-scale model initially gives acceptable residuals on the team's calibration images.

When operators bring the component closer and tilt it more strongly, the residuals increase systematically across its depth. The team checks the viewing geometry and compares a perspective model instead of assuming that every mismatch is a detector failure.

The case illustrates how a convenient geometric approximation can have a restricted operating range. No accuracy improvement or measured deployment result is claimed.

Common misconceptions and mistakes

Weak perspective uses no depth information at all. Its shared scale depends on representative depth.
It is identical to full perspective. It removes individual point-depth scaling.
Small objects always justify the approximation. Their distance and required tolerance also matter.
Changing orientation has no effect. Rotation changes the coordinates being projected.
A shared scale fixes lens distortion. Distortion requires its own model or correction.
A low fitting error proves a unique three-dimensional reconstruction. Projection can leave depth and scale ambiguities.

Business use

For business applications, the relevant question is whether the approximation is accurate enough over the expected viewing conditions. State the assumed distance range, depth variation and tolerated feature error.

Test near and far placements, rotations and positions across the image rather than evaluating only a convenient central view. If a downstream decision depends on millimeters, establish how image error translates into measurement uncertainty.

A simpler camera model can make an estimation problem easier to handle, but savings in computation do not establish geometric validity. Retain a comparison against calibrated perspective predictions where the distinction affects the product's output.

Check your understanding

With focal length 600 pixels and representative depth 12 meters, what shared scale is used, and where does a point with X=2 meters project horizontally relative to the principal point?

Answer: The scale is 600/12=50 pixels per meter, giving a horizontal offset of 100 pixels under the scaled orthographic model.

Related terms

Perspective projection
Orthographic projection
Structure from motion
Weak perspective
camera calibration
focal length
principal point
pose estimation

Learn next

Calculate perspective and weak-perspective offsets for one point while varying its depth around a fixed representative depth. Plot their difference and choose a pixel-error tolerance. Identify which depths meet that tolerance before using the approximation.

Sources

Subhashis Banerjee's camera-model notes, with weak-perspective equations and approximation conditions: https://www.cse.iitd.ac.in/~suban/vision/geometry/node29.html

The relevant equations and explanatory text were reviewed. The numerical comparison and Maple Vision case are original illustrations, not empirical results from the source.

Previous: Scale-space axioms Next: Scales of measurement