Does a Euclidean transform change size?

No. It preserves distances, so an object's geometric size remains unchanged.

Can it include reflection?

In general mathematical usage, Euclidean isometries can include reflections. Many vision tools use the term for orientation-preserving rotation and translation only.

Is every affine transform Euclidean?

No. Affine transformations can include scaling and shear, which do not generally preserve distances.

AI Glossary / Computer vision and multimodal AI / Beginner

Euclidean transformation

Definition

A Euclidean transformation preserves distances and angles while changing the position or orientation of a geometric object. In common computer-vision usage, it means a rigid transformation made from rotation and translation, without scaling or shearing. In broader mathematics, distance-preserving transformations can also include reflections, so the intended convention should be stated.

What it means

Imagine sliding and turning a rigid shape on a table. Its corners change coordinates, but the distances between them stay the same. That is the central idea behind a Euclidean or rigid transformation in vision.

The transformation changes how points are located in a coordinate system, not the intrinsic size or shape of the object. It is therefore useful when two observations differ mainly by position and orientation.

A real camera image may require a more flexible model. Perspective changes can alter apparent size and shape even when the physical object is rigid. A two-dimensional Euclidean transform is appropriate only when its assumptions fit the relationship between the images being aligned.

How it works

Represent each point in a stated coordinate system. In the rotation-and-translation convention, the transformed point is q = Rp + t, where R is a rotation matrix and t is a translation vector.

Choose or estimate the rotation and translation from suitable observations. In two dimensions, the rotation can be described by one angle. In three dimensions, orientation has three degrees of freedom, with several possible mathematical representations.

Apply the same transformation consistently to the relevant points. For an image warp, determine how output locations map back to input locations and use an interpolation method to obtain pixel values.

Check residual differences against known correspondences. If the required alignment involves scale changes, shear, perspective, or deformation, a rigid model may leave systematic error. Increasing optimization effort cannot make an unsuitable transformation family represent changes it excludes.

Three practical examples

1. Planar alignment: A scan-processing tool corrects a shifted and slightly rotated document when scale and perspective differences are negligible.

2. Point-cloud registration: A system aligns observations of a rigid object using a three-dimensional rotation and translation.

3. Coordinate conversion: A robot expresses a measured point in another rigidly related coordinate frame using a known transformation.

Think of it this way

Think of moving a transparent stencil over a drawing. You can slide and turn the stencil while its printed shape remains unchanged. Stretching or skewing it would require a different transformation.

The analogy is limited because image resampling can change pixel values even when the underlying coordinate transform is rigid. Also, a physical object turning in three dimensions can produce a perspective change that a simple two-dimensional stencil motion does not capture.

Worked example

Take two points A = (1, 0) and B = (4, 0). Their distance is 3 units. Rotate them 90 degrees counterclockwise about the origin, using ordinary Cartesian coordinates with the vertical axis pointing upward.

The rotated points are A1 = (0, 1) and B1 = (0, 4). Now translate both by (5, 2), producing A2 = (5, 3) and B2 = (5, 6). Their distance remains 3.

Order matters when rotation and translation are specified relative to fixed coordinates. If A is translated first, it becomes (6, 2); rotating that point 90 degrees gives (-2, 6), which differs from (5, 3).

Both operation sequences preserve distances, but they describe different final placements. In an image array whose row coordinate increases downward, the visual interpretation of a positive angle may differ from this Cartesian example, so software conventions must be checked before comparing results.

Illustrative case study

Fictional case study: ScanFit aligns repeated scans of a printed reference card. A Euclidean transform corrects most position and rotation differences, but one scan still shows increasing error toward the corners.

The team discovers that the card was photographed at an angle rather than scanned flat. It compares a more suitable projective model and reviews whether the available correspondences support that additional flexibility.

It keeps rigid alignment for cases where the simpler assumptions hold. This fictional case illustrates model selection; it does not claim that the most flexible transform is always best or that a rigid physical object always produces rigidly related images.

Common misconceptions and mistakes

Do not include arbitrary scaling in a rigid Euclidean transform. Rotation, translation, and uniform scaling together belong to a similarity-transform model in common vision terminology.

Do not confuse physical rigidity with a two-dimensional rigid image relationship. Camera projection can introduce apparent changes that require another model.

Do not ignore operation order or frame direction. A transform from frame A to frame B is not the same as its inverse.

Do not assume distance preservation in geometry means pixel-perfect preservation after warping. Interpolation, cropping, and finite sampling can alter the raster image.

Business use

Euclidean transformations are valuable when alignment should preserve shape and scale. Their limited number of parameters can make them interpretable and easier to estimate than more flexible models when the assumptions are appropriate.

Specify coordinate units, axis directions, transform direction, and the permitted operations. Evaluate errors across the whole object or image rather than only near the fitted points. If residuals show systematic scaling or perspective effects, review the model instead of treating every mismatch as random noise or a failed optimizer.

Check your understanding

A transformation doubles every distance while preserving angles. Is it a rigid Euclidean transform under the common vision convention?

Answer: No. The scale change violates distance preservation. It fits a similarity-transform description if the remaining operations are rotation and translation.

Related terms

Affine transformation
Image registration
Rigid transformation
rotation matrix
translation
similarity transformation
coordinate frame

Learn next

Learn similarity and affine transformations next to see which geometric properties each preserves. Then practice composing and inverting rigid transforms, keeping coordinate frames and operation order explicit.

Sources

scikit-image, EuclideanTransform in the transform API: primary software documentation consulted for the rotation-and-translation convention used in image processing. The broader mathematical reflection distinction is stated separately. The point calculation and fictional scan case are original teaching examples. https://scikit-image.org/docs/stable/api/skimage.transform.html

Previous: Euclidean space Next: Euler angle