Why are triangles convenient?

Three noncollinear correspondences determine a two-dimensional affine map, and barycentric coordinates provide a direct way to map interior points.

Does the warp preserve straight lines everywhere?

Within each triangle, affine mapping preserves straightness. A line crossing multiple triangles can bend at their boundaries.

Is adding more control points always beneficial?

No. Poorly placed points can create narrow, unstable, or overly distorted triangles and make the deformation harder to manage.

AI Glossary / Computer vision and multimodal AI / Intermediate

Mesh / triangulation warp

Definition

A mesh or triangulation warp deforms an image using a network of control points connected into triangles. In a common piecewise-affine version, each triangle receives the affine transformation determined by its three source and destination vertices. The combined mapping provides flexible local deformation while keeping shared edges connected.

What it means

A single affine transformation moves an entire image according to one rule. That is useful for uniform translation, scaling, rotation, or shear, but it cannot independently adjust several local regions.

A triangulated warp divides the image into smaller areas. Moving selected vertices changes the triangles around them, allowing different parts of the image to stretch or shift differently. The mapping is exact at the specified control points when the triangles and correspondences are valid.

This entry concerns a two-dimensional image warp, not every deformation of a three-dimensional mesh. A triangulation supplies the spatial organization; the actual image colors still need to be sampled appropriately after the geometry changes.

How it works

Choose corresponding source and destination control points and connect them with a consistent triangulation. Three noncollinear vertex correspondences determine an affine mapping for each triangle.

For a point inside a triangle, calculate its barycentric coordinates: three weights summing to one that describe its position relative to the vertices. Apply the same weights to the corresponding destination vertices to locate the mapped point.

To generate the output image, inverse mapping is commonly used: find each destination pixel's source position and interpolate its color. Shared vertices and edges make the piecewise mapping continuous when assembled consistently, but derivatives can jump at triangle boundaries. Check for degenerate or flipped triangles, overlapping regions, missing coverage, and sampling artifacts before treating the warp as usable.

Three practical examples

1. Image morphing: A designer marks corresponding features in two images and uses a common triangle layout to interpolate their positions. Color blending is a separate step from the geometric warp.

2. Document correction: A prototype uses control points to adjust a distorted page locally. It evaluates whether the resulting text and line geometry remain faithful, particularly near triangle boundaries.

3. Texture placement: An artist maps a flat pattern into a specified two-dimensional region. The triangle mesh provides local control while image resampling determines the appearance of the mapped texture.

Think of it this way

Think of a sheet divided into triangular panels whose corners can be repositioned. Each panel stretches according to its three corners, and adjacent panels share their common edge.

The analogy does not imply physically realistic cloth. A piecewise-affine warp does not automatically enforce material stiffness, prevent overlap, or preserve area, and its change in slope between panels can produce visible creases in the deformation.

Worked example

Take a source triangle with vertices A = (0, 0), B = (4, 0), and C = (0, 4). The point P = (1, 1) has barycentric weights 0.5 for A, 0.25 for B, and 0.25 for C because their weighted sum is (1, 1).

Move the vertices to A′ = (0, 0), B′ = (8, 0), and C′ = (0, 4). Applying the same weights gives P′ = 0.5A′ + 0.25B′ + 0.25C′ = (2, 1). This triangle has doubled horizontally while keeping its vertical scale.

The midpoint of edge BC has weights (0, 0.5, 0.5), so it maps to (4, 2). An adjacent triangle sharing the same mapped B and C sends that edge point to the same location, supporting continuity across the shared edge.

However, the neighboring triangle can have a different interior stretch. Agreement on the edge therefore does not guarantee matching derivatives across it. The arithmetic locates points; interpolation and antialiasing are still needed to render pixel values.

Illustrative case study

Fictional case: A design agency warps a photographed package label to fit a revised outline. The control points align correctly, but straight decorative lines bend abruptly where they cross several triangle edges.

The artist reviews the triangulation and adjusts the control layout, then compares the result with a smoother deformation method. The team also checks small text at final output size because repeated resampling can reduce its clarity.

The scenario illustrates the difference between satisfying point correspondences and achieving the desired overall deformation. It does not report a real production result or guarantee that more triangles always improve the image.

Common misconceptions and mistakes

Do not confuse geometric continuity with smoothness of every derivative. A standard piecewise-affine mesh is generally continuous but can have changes in slope across triangle boundaries.

Do not assume valid source triangles remain valid after their vertices move. Collinear vertices create a degenerate mapping, while flipped or overlapping triangles can make the intended inverse mapping ambiguous.

Avoid treating triangulation and resampling as the same operation. A correct coordinate map can still produce jagged edges, blur, or aliasing if pixels are sampled poorly. Also distinguish exact control-point fitting from physical accuracy: incorrect correspondences can be fitted perfectly.

Business use

Triangulation warps offer interpretable local control for editing, alignment, and shape-based image transformations. They can be attractive when a modest number of meaningful correspondences describes the desired deformation.

Evaluate control-point placement, triangle quality, continuity, and local distortion separately from visual texture quality. Keep the original image and control mesh so edits remain reproducible. If the warp supports measurement rather than presentation, quantify its geometric error against independent references and document regions where extrapolation or weak control makes the result uncertain.

Check your understanding

Two neighboring triangles agree on their shared edge. Must they have the same stretch on either side?

Answer: No. Their mappings can be continuous along the edge while having different derivatives in their interiors.

Related terms

Affine transformation
Barycentric coordinates
Image morphing
Inverse mapping
triangulation
control point
interpolation
thin-plate spline

Learn next

Map one point by barycentric coordinates, then join two triangles and move their outer vertices independently. Inspect shared-edge continuity and changes in local stretch before adding image resampling to the experiment.

Barycentric coordinates

Sources

Frédo Durand, Morphing, section 6.3.5 Mesh / triangulation morphing: https://people.csail.mit.edu/fredo/comp-photo-book/06-warping-and-morphing-03-morphing.html

The author's technical chapter was opened, including its piecewise-affine triangulation discussion. The coordinate calculation and label-design scenario are original illustrations.

Previous: Mesh subdivision Next: Mesokurtic