Motivation
A cross-dissolve between two portraits changes opacity while leaving both sets of eyes in their original positions. Halfway through, the viewer may see two faint pairs of eyes. A geometric morph instead asks which point in one image should become which point in the other, moves those features toward a common intermediate shape, and then combines appearance at corresponding locations. Geometry and color interpolation are separate decisions. Neither becomes correct merely because the transition looks smooth in time.
The hard part is often the correspondence. A point selected on the left corner of one mouth should have the same semantic role in the second image. Unmatched occlusions, different expressions and a newly visible surface cannot always be described by a one-to-one planar warp. We begin with a finite model where corresponding landmarks and a common triangle connectivity are supplied. This makes the numerical construction explicit and lets us inspect exactly where an inverse mapping exists.
Even this simple model contains a surprise: nondegenerate, consistently oriented endpoint triangles can collapse at an intermediate time. Checking the first and last frames is insufficient. We will distinguish continuity across triangle edges, local invertibility inside a triangle, and global one-to-one behavior across the whole mesh. These properties answer different questions and need different hypotheses.
Definitions
Definition 1.28.1
(Conforming common triangulation and interpolated landmarks)
Let a finite conforming triangulation cover a compact convex polygonal domain . Conforming means that distinct triangles have disjoint interiors and intersect only in a whole common edge, a common vertex, or not at all; there are no hanging vertices. Globally indexed vertices describe the first image and corresponding vertices the second. The same triples of indices define triangles in both meshes. For time , interpolate landmarks by . A chosen intermediate triangle must have nonzero area before it can be used for inverse sampling.
Definition 1.28.2
(Barycentric coordinates and the triangle affine map)
For an ordered source triangle , write
If , the barycentric coordinates are unique. A point lies in the closed triangle exactly when all three coordinates are nonnegative. On an edge, the coordinate of the opposite vertex is zero. The affine map to the corresponding target triangle is
Its Jacobian determinant is . Zero means local collapse; a negative value reverses the ordered orientation. A foldover may involve such a reversal or overlap between different mesh regions. Positive orientation in each individual triangle does not, by itself, constitute a global non-overlap proof for an arbitrary assembled mesh.
Definition 1.28.3
(Two-image morph value at an output point)
To produce an intermediate image at a point in triangle , find its barycentric coordinates and query
Here is a specified continuous reconstruction rule, such as bilinear interpolation. This is a geometric warp followed by cross-dissolve in a common intermediate geometry. Linear-light appearance values give a physical mixture interpretation; display-encoded blending generally gives a different result.
Theorem/proposition
Theorem 1.28.4
(Vertex correspondence gives a unique continuous piecewise affine map)
On the conforming source triangulation above, suppose every source triangle is nondegenerate and target values are assigned consistently to globally indexed vertices. There is a unique map that sends each to and is affine on every source triangle. It is continuous on the triangulated domain. On a triangle it is invertible exactly when the corresponding target triangle is nondegenerate. Continuity does not imply global injectivity.
Theorem 1.28.5
(A small Lipschitz displacement gives a global inverse on the image)
Let be a continuous displacement satisfying for all . If , the map satisfies
It is injective and has an inverse on with Lipschitz constant at most . This is a sufficient condition, not a necessary condition for a valid morph.
For a continuous piecewise affine displacement on a convex triangulated domain, a usable bound is the largest Euclidean operator norm of its triangle Jacobians. A line segment between two domain points can be split at triangle boundaries, and summing the bounds along its pieces gives the global Lipschitz estimate. Convexity ensures that the segment stays in the domain. On a nonconvex domain, that particular segment argument needs modification.
Proof sketch or proof idea
Proof
(Existence, continuity and local invertibility of the piecewise affine map)
On a nondegenerate source triangle put . It is affine, and since and it sends to . It is the only affine map with these vertex values: the difference of two such maps is affine, vanishes at , and so is a linear map that kills the columns of the invertible matrix , hence is zero.
Two triangles of a conforming triangulation meet in a whole common edge, a common vertex, or not at all. On a common edge with endpoints , a point is , and either triangle's affine map sends it to , because both use the same globally indexed targets. The pieces therefore agree wherever they overlap; finitely many closed pieces on which the map is continuous, agreeing on overlaps, give a continuous map on the whole domain. On a triangle, the linear part is , which is invertible exactly when , that is, when the target triangle is nondegenerate. None of this rules out two distant triangles mapping to overlapping regions.
Proof
(Distance bounds from a small Lipschitz constant)
For ,
The triangle inequality and give . The reverse triangle inequality gives . When the constant is positive, so forces : the map is injective and has an inverse on its image. For image points and the lower bound reads , which is the inverse Lipschitz bound. As approaches one, the guarantee weakens: a small output displacement could correspond to a large source displacement.
Proof X-ray
Shared edges give continuity; a lower distance bound gives injectivity
The first proof uses only agreement of endpoint data along an edge. It cannot control where another triangle goes. The second proof compares every pair of points and prevents their images from colliding. These are different logical mechanisms. A mesh renderer can hide an overlap by choosing whichever triangle was drawn last, but that drawing convention does not turn a many-to-one geometric map into an inverse.
Each hypothesis enters at one place. Consistent global indexing is used only to make the two edge formulas equal. Nondegeneracy of is used only to form . The condition is used only to make positive, which turns a lower distance bound into injectivity. Convexity of is not used in either proof; it enters only in the segment argument that bounds for a piecewise affine displacement.
There is also an exact finite check for the area of each interpolated triangle. Its edge matrix is , so is a quadratic polynomial in two dimensions. To prove throughout an interval, inspect its endpoints and any interior critical point that can be a minimum. Merely observing positive values at a few frames may miss a root between them. A nonzero area check is local; global overlap and boundary behavior still need a separate test or a sufficient theorem such as the distance bound above.
Worked examples
Worked example 1.28.6
(An entire intermediate pixel calculation)
Take source vertices and target vertices . At , intermediate vertices are . The output point has barycentric coordinates . Its source queries are and . For continuous images and , the sampled values are and , so the intermediate value is . These are illustrative intensity units, not a claim that a stored channel is bounded by one.
Worked example 1.28.7
(Both endpoints are oriented correctly, but the path collapses)
Let and . Both determinants equal one. However,
It vanishes at and , and is negative between them. At its value is . A half-turn together with unequal scaling can therefore pass through collapsed triangles under straight vertex interpolation. The remedy requires a different geometric path, additional correspondences or a controlled deformation model; a different color blend cannot restore an absent geometric inverse.
Worked example 1.28.8
(A conservative global guarantee)
Suppose with . Its Euclidean operator norm is . For every , the theorem gives , so no pair collides. At , the actual linear part is , whose smallest singular value is ; the bound is exact in the vertical direction. The translation changes position but not the difference bound. A large pure translation is therefore harmless even though absolute landmark displacement is large.
For a digital implementation, first choose corresponding landmarks and one connectivity, then verify source and target triangles and the intended time interval. For each output frame, locate each output pixel in a valid intermediate triangle, compute barycentric coordinates, query both original images, and blend. Shared edges need a consistent ownership convention so that a pixel on an edge is neither missed nor written inconsistently. The continuous formulas agree there, but floating-point triangle tests can still disagree unless the rasterization rule is explicit.
Concept lensAlgorithmic
A morph samples two original images through one moving mesh
At an intermediate location, the barycentric coordinates select corresponding source points in both originals. The intermediate frame is not obtained by repeatedly warping the previous frame. Sampling originals avoids accumulating interpolation blur across time. When a triangle shrinks, its inverse map covers a large source footprint, so antialiasing may still be needed. Avoiding repeated resampling does not remove the sampling theorem's information limits.
Connectivity deserves its own check. Computing an unrelated triangulation in each endpoint may choose different diagonals of a quadrilateral. Then the phrase “corresponding triangle” is undefined on part of the image. A common combinatorial mesh is part of the model, not a detail to leave to two independent triangulation calls. Landmarks on the image boundary also matter: without them, an intended full-frame morph may leave uncovered output regions.
Common mistakes
Counterexample mode
Continuity does not prevent overlap
On the one-dimensional interval , the continuous piecewise affine function agrees at its shared vertex zero, yet maps and to the same point. Extending it as on a rectangle produces an overlapping planar map. The local formulas fit together without a crack, but one side has reversed orientation and the whole map has no single-valued inverse. Edge agreement alone cannot certify a usable image warp.
Common mistake
(Treating every landmark as the same visible material point)
When feature visibility changes, a landmark pair may not label one material point. A closed mouth and an open mouth show different visible surfaces; no planar bijection can invent the teeth hidden in the first image. Cross-dissolving may make new appearance emerge, but that is a synthesis choice. Similarly, moving the boundary correctly does not guarantee that texture inside each large triangle follows a plausible motion.
Common mistake
(Reading a positive determinant as numerical safety)
Invertibility is a yes/no statement; numerical sensitivity depends on singular values. A skinny triangle may amplify a small landmark error even when its area is positive. Inspect condition numbers, meaningful correspondence density and source footprints rather than relying only on determinant sign. More landmarks are useful when they represent information, not when they merely subdivide an unsupported guess.
Summary
A piecewise affine morph uses shared landmark connectivity, barycentric geometry and appearance interpolation. Matching edge values gives continuity, while nonzero triangle area gives a local inverse. Neither alone guarantees global one-to-one behavior. A Lipschitz displacement bound supplies one rigorous global sufficient condition, and the quadratic area polynomial exposes intermediate collapses missed by endpoint checks. Valid geometry must still be combined with careful sampling and explicit treatment of appearance and occlusion.
Exercises
Quick check
(In the first worked example, take intermediate point at . Find its barycentric coordinates, both source queries and the final intensity for the same two continuous images.)
Solution · Answer
The barycentric coordinates are because and . Thus and . The intensities are and , and their half blend is one. The same barycentric weights are used for geometry in both images; using the intermediate Cartesian coordinate directly in either original would query a different feature.
Quick check
(Two adjacent source triangles share vertices . Prove directly that their affine maps agree at for , provided the target values at those vertices agree. What fails if one triangle uses a different target for ?)
Solution · Answer
Affinity gives from either triangle, so the restrictions agree throughout the edge. If the second triangle uses , the edge values differ by , nonzero for . The assembled map then has a crack along the edge rather than merely a numerical interpolation discrepancy.
Quick check
(For and , expand and locate its minimum on . Verify that checking just and misses the collapse.)
Solution · Answer
Expansion gives . Its derivative vanishes at , where . At both endpoints . Its roots are and , so the endpoint signs conceal both singular times and an orientation-reversing interval.
Quick check
(A continuous piecewise affine displacement on a convex domain has triangle Jacobian norms at most . Give lower and upper distance bounds and an inverse Lipschitz bound at . Explain why the theorem does not require a small constant translation component.)
Solution · Answer
Here , hence , and the inverse constant is at most . A constant displacement cancels in and contributes zero to every Jacobian. The condition controls relative deformation, not how far the entire image translates.