6. Curves, Surfaces, and Mesh Processing

Introduction and Context

Computer graphics represents shape at several levels. A polygon mesh provides an explicit discrete approximation, whereas a curve or surface provides a continuous mathematical description from which geometry can be sampled. Curves are central to fonts, paths, animation trajectories, industrial design, and camera motion. Surfaces define the form of manufactured objects, characters, terrain, and scientific models. Mesh processing then converts, edits, analyzes, and optimizes these representations for rendering and simulation.

The central engineering problem is not merely to create a visually plausible shape. A representation must also be controllable, numerically stable, sufficiently accurate, efficient to store and render, and suitable for downstream operations such as tessellation, collision detection, deformation, and physical simulation. This chapter develops the mathematical foundations of Bézier, Hermite, B-spline, and NURBS representations and connects them with surface construction and mesh-processing operations.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the geometric and algebraic properties of common parametric curves.

2. Construct and evaluate Bézier, Hermite, B-spline, and NURBS curves.

3. Extend curve representations to tensor-product, swept, and subdivision surfaces.

4. Select tessellation strategies according to curvature, error, and performance requirements.

5. Apply parameterization, smoothing, simplification, deformation, and quality assessment methods.

6. Analyze the numerical, visual, and production trade-offs of alternative representations.

7. Identify common geometric-processing failures and propose appropriate remedies.

A Mathematical View of Shape

A parametric curve is a mapping

p(t) = (x(t), y(t), z(t)), t in [a,b].

Unlike an implicit curve, which is described by an equation such as f(x,y)=0, a parametric curve directly supplies a point for each parameter value. The derivative p'(t) gives a tangent direction, and the second derivative helps characterize curvature. If p'(t) is nonzero, the unit tangent is

T(t) = p'(t) / ||p'(t)||.

For a planar curve, curvature may be written as

kappa(t) = |x'(t)y''(t)-y'(t)x''(t)| / (x'(t)^2+y'(t)^2)^(3/2).

High curvature generally requires denser sampling during tessellation. However, curvature alone is not always sufficient: a small geometric feature, a sharp change in normal, or a texture coordinate discontinuity may also require refinement.

A surface is a mapping s(u,v) from a two-dimensional parameter domain to three-dimensional space. Its tangent vectors are s_u and s_v, and the surface normal is proportional to s_u cross s_v. A valid regular surface requires these tangent vectors to be linearly independent. Degenerate parameterizations can produce collapsed triangles, unstable normals, and poor numerical behavior.

Bézier Curves: Global Control with a Clear Geometric Interpretation

An nth-degree Bézier curve defined by control points P0,...,Pn is

B(t) = sum from i=0 to n of B_i,n(t) P_i, 0 <= t <= 1,

where

B_i,n(t) = C(n,i)(1-t)^(n-i)t^i

are Bernstein polynomials. The basis functions are nonnegative and sum to one. Consequently, the curve lies inside the convex hull of its control points. This property is valuable for intuitive design and geometric bounding.

The endpoint conditions are B(0)=P0 and B(1)=Pn. Endpoint tangents are

B'(0)=n(P1-P0), B'(1)=n(Pn-Pn-1).

Thus, the first and last control edges determine the endpoint directions. A quadratic curve uses three control points; a cubic curve uses four and is the most common practical form because it provides useful shape control without excessive degree.

The de Casteljau algorithm evaluates a Bézier curve through repeated linear interpolation. For parameter t, intermediate points are computed as

P_i^(1)=(1-t)P_i+tP_i+1,

and the process continues until one point remains. Besides being numerically stable, the algorithm subdivides a Bézier curve into two Bézier curves. This makes it useful for adaptive tessellation, intersection algorithms, and bounding-volume construction.

A major limitation of a single high-degree Bézier curve is global control: moving one control point can affect the entire curve. High degree may also cause numerical sensitivity and unwanted oscillation. Complex shapes are therefore usually represented by piecewise Bézier segments joined with continuity constraints.

Hermite Curves: Interpolation Through Geometric Conditions

A cubic Hermite curve is specified by two endpoints P0 and P1 and two endpoint tangent vectors T0 and T1:

H(t) = h00(t)P0 + h10(t)T0 + h01(t)P1 + h11(t)T1,

where

h00=2t^3-3t^2+1,

h10=t^3-2t^2+t,

h01=-2t^3+3t^2,

h11=t^3-t^2.

It satisfies H(0)=P0, H(1)=P1, H'(0)=T0, and H'(1)=T1. Hermite form is particularly appropriate when interpolation data and derivatives are known, as in camera paths, motion planning, and physical trajectories.

The distinction between Bézier and Hermite form is often one of representation rather than expressive power. A cubic Hermite segment can be converted to a cubic Bézier segment with control points

Q0=P0,

Q1=P0+T0/3,

Q2=P1-T1/3,

Q3=P1.

This conversion allows systems to use Hermite-style input while retaining Bézier evaluation and subdivision procedures.

When multiple segments are joined, continuity must be considered. C0 continuity means positions coincide. C1 continuity means tangent vectors coincide, and G1 continuity means tangent directions coincide, possibly with different magnitudes. C2 continuity additionally matches curvature-related behavior. In visual design, G1 may be sufficient; in machining or high-quality animation, stronger continuity can be important.

B-Splines: Local Control and Flexible Degree

A B-spline curve is defined by control points Pi, degree p, and a nondecreasing knot vector U={u0,...,um}:

C(u)=sum Ni,p(u)Pi.

The basis functions are defined recursively by the Cox–de Boor relation. At degree zero, Ni,0(u) is one over its knot interval and zero elsewhere. For p greater than zero,

Ni,p(u)=((u-ui)/(ui+p-ui))Ni,p-1(u)+((ui+p+1-u)/(ui+p+1-ui+1))Ni+1,p-1(u),

with terms having zero denominators interpreted as zero.

At any parameter value, only a limited number of basis functions are nonzero. This local support means that changing one control point affects only a portion of the curve, unlike a single high-degree Bézier curve. Degree, knot spacing, and endpoint treatment jointly determine the shape.

A clamped knot vector repeats the first and last knots p+1 times, causing the curve to interpolate the first and last control points. Uniform knot vectors provide regular parameter intervals, while nonuniform vectors allow local refinement and varying influence. Knot insertion changes the representation without changing the curve geometry. This is important for editing, adaptive refinement, and conversion to piecewise Bézier form.

B-splines generally do not interpolate interior control points. This is an advantage for smooth design but can surprise users expecting control points to lie on the curve. Interpolation can be imposed by solving for suitable control points, although this may reduce locality or require additional constraints.

Rational Curves and NURBS

Polynomial curves cannot represent exact circles and many other conic sections. Rational curves overcome this limitation by assigning weights wi to control points:

R(u)= [sum Ni,p(u) wi Pi] / [sum Ni,p(u) wi].

A Non-Uniform Rational B-Spline, or NURBS, combines B-spline basis functions with weights and a nonuniform knot vector. In homogeneous coordinates, each control point is represented as (wi xi, wi yi, wi zi, wi), and the final three-dimensional point is obtained by perspective division by the fourth coordinate.

Increasing a control-point weight pulls the curve toward that point, but weights must be interpreted carefully because their effect is relative to neighboring weights. NURBS support exact lines, circles, ellipses, and many engineering forms, which explains their importance in CAD, manufacturing, and exchange standards.

NURBS also introduce concerns. Poor weights can create unexpected shapes, and parameter-space uniformity does not imply uniform spatial speed. Robust implementations must handle knot multiplicities, zero or near-zero denominators, domain conventions, and conversion between modeling and rendering coordinate systems.

From Curves to Surfaces

A tensor-product Bézier surface is defined by a rectangular control net:

S(u,v)=sum_i sum_j B_i,m(u)B_j,n(v)Pij.

The surface lies within the convex hull of its control net and interpolates corner control points. Tensor-product B-spline and NURBS surfaces replace Bernstein bases with B-spline or rational bases. Their local support provides practical control for large surface models.

Other constructions are useful for different design situations. A ruled surface is generated between corresponding points on two curves. A loft surface interpolates or approximates a sequence of profile curves. A sweep moves a cross-section along a spine, requiring a stable moving frame to avoid twisting. A surface of revolution rotates a profile around an axis. Each construction encodes a design intention, and the choice affects parameterization, continuity, and downstream processing.

A mesh is a discrete surface composed of vertices, edges, and faces. Its topology records adjacency, while geometry records vertex positions. Triangle meshes are widely used because graphics hardware and many numerical algorithms operate naturally on triangles. Quad meshes may be preferable for subdivision, deformation, and character animation because their edge flow can follow semantic features.

Subdivision surfaces refine a coarse control mesh and apply a limit rule. Catmull–Clark subdivision is commonly associated with quadrilateral control meshes, while Loop subdivision is designed for triangular meshes. Subdivision offers smooth multiresolution modeling, but extraordinary vertices and boundary rules require careful handling. The limit surface may be evaluated directly or approximated by tessellated patches.

Tessellation: Turning Continuous Geometry into Renderable Meshes

Tessellation samples a curve or surface and generates line segments or polygons. Uniform sampling is simple but inefficient: it oversamples flat regions and undersamples high-curvature regions. Adaptive tessellation instead subdivides according to a geometric error criterion, such as the distance between a curve and its chord or the deviation of a surface patch from a planar approximation.

For a surface patch, refinement may be triggered by one or more of the following:

• positional deviation from the patch’s approximation;
• normal variation or curvature;
• projected screen-space size;
• texture or displacement detail;
• silhouette proximity;
• cracks between neighboring patches.

Screen-space adaptive tessellation is efficient for real-time rendering because it allocates triangles according to visual importance. Object-space tolerances are more appropriate when the mesh will be manufactured, simulated, or reused from different viewpoints. A robust tessellator also preserves consistent shared boundaries; otherwise, independently refined patches can produce visible T-junctions and cracks.

Degenerate triangles, inverted parameter orientation, and excessively thin triangles reduce rendering and simulation quality. Tessellation should therefore be treated as a constrained approximation problem rather than simple point sampling.

Parameterization and Mesh Quality

A parameterization assigns a two-dimensional domain coordinate to each surface point or mesh vertex. Common methods include uniform, chord-length, centripetal, and arc-length parameterization for curves. Uniform parameters are easy to compute but may produce uneven spatial spacing. Chord-length parameters reflect distances between samples, while centripetal parameterization often reduces undesirable cusps and loops in interpolating splines.

For triangular meshes, quality is assessed using several complementary measures. The aspect ratio compares the longest edge with an appropriate altitude or shortest edge. A triangle with a very small angle is usually undesirable. The minimum angle, maximum angle, area, and radius ratio are common indicators. For numerical simulation, element conditioning and alignment with physical fields may matter more than visual appearance.

Surface parameterization can be global or local. Planar projection is fast but causes distortion when the surface is not nearly planar. Angle-based methods attempt to preserve angles, whereas area-based methods attempt to preserve area. No parameterization preserves all desirable properties for arbitrary surfaces. Closed surfaces and surfaces with complex topology require cuts, charts, or specialized global methods.

Texture mapping makes parameterization visibly important. Stretching changes texture scale, while angular distortion changes shape. A production pipeline should inspect distortion heat maps and seams rather than relying only on the rendered texture.

Mesh Smoothing and Fairing

Smoothing removes noise or improves visual regularity. The simplest method is Laplacian smoothing:

vi_new = (1-lambda)vi + lambda average(Ni),

where Ni is the one-ring neighborhood. Repeated application can reduce high-frequency noise, but it also causes shrinkage, moves boundaries, and may destroy sharp features.

Taubin smoothing alternates positive and negative filtering steps to reduce shrinkage while retaining the low-pass character of the operation. For feature-sensitive processing, weights may depend on edge length, curvature, or normal difference. Bilateral and robust methods smooth positions while attempting to preserve discontinuities.

Smoothing should distinguish noise from intentional detail. A crease, corner, or engraved feature may be geometrically important even if it resembles high-frequency variation. Boundary constraints, marked feature edges, and volume-preserving objectives should be incorporated when appropriate.

Simplification and Level of Detail

Mesh simplification reduces polygon count while attempting to preserve appearance, topology, and important features. A widely used strategy is edge collapse: an edge joining vertices vi and vj is replaced by a single vertex, and adjacent faces are updated. The operation is prioritized by a cost function.

Quadric error metrics approximate the squared distance from a candidate vertex to the planes of affected faces. The accumulated quadric provides an efficient estimate of geometric deviation. A simplifier may also penalize boundary movement, normal changes, texture distortion, material-boundary crossings, or topology changes.

A production system commonly generates a level-of-detail hierarchy rather than one simplified mesh. LOD selection may depend on projected size, distance, motion, platform capability, and frame-time budget. Aggressive simplification can remove thin parts, merge separate features, create holes, or produce shading artifacts. Visual inspection and automated geometric checks should therefore operate together.

Deformation and Editing

Deformation changes shape while attempting to preserve continuity, volume, or semantic structure. Direct vertex manipulation is simple but often produces local artifacts. Skeletal deformation uses a joint hierarchy and skinning weights; linear blend skinning is efficient but can collapse volume near joints. Dual-quaternion skinning better preserves rotational volume in many articulated situations.

Free-form deformation embeds an object in a lattice and maps lattice control points to deformed positions. It provides smooth global control but may affect regions unintentionally. Differential-coordinate and Laplacian-based methods preserve local detail by maintaining relationships among neighboring vertices. Cage-based deformation offers an intuitive boundary-driven alternative.

For editing CAD-like surfaces, continuity constraints and feature curves are essential. For animation, deformation must also respect rigging, collision, temporal coherence, and shading normals. The appropriate method depends on whether the priority is local detail preservation, global shape change, physical plausibility, or interactive speed.

Applied Scenario: Preparing a Scanned Mechanical Housing for Real-Time Use

Background

Consider an illustrative industrial visualization project in which a mechanical housing is acquired through 3D scanning. The scan contains millions of noisy points, small holes, uneven density, and sharp machined edges. The final asset must support inspection on a real-time platform and must also permit accurate measurement.

Situation and Analysis

A direct conversion of the point cloud into a uniformly dense mesh would be expensive and would blur important edges. The team first identifies the required tolerances. Measurement regions require object-space accuracy, whereas distant visualization regions can use lower-detail representations. Sharp edges are marked as semantic features rather than treated as noise.

Approach

A screened reconstruction produces an initial watertight surface, after which holes are classified as either scanning defects or intentional openings. The surface is repaired using feature-aware operations. An adaptive remeshing stage improves triangle regularity while preserving marked boundaries. The high-resolution model is retained as a reference.

For interactive use, a quadric-error simplifier generates several LODs with penalties for boundary displacement and material transitions. Normals are recomputed with hard-edge rules. Tessellation is screen-aware for curved regions, but inspection mode uses a stricter object-space tolerance. Texture coordinates are generated with a distortion diagnostic, and automated tests check for nonmanifold edges, inverted faces, zero-area triangles, and excessive aspect ratios.

Outcome

The resulting asset provides accurate close inspection and efficient distant rendering. The central design decision is not a single algorithm but a pipeline in which representation, tolerance, topology, and visual purpose are coordinated.

Learning Points

The scenario demonstrates that mesh processing is application-dependent. Noise removal must not erase designed features; simplification must account for semantics; and quality assessment must include both geometric and rendering criteria. A reference model, explicit tolerances, and multiple LODs make the pipeline measurable and maintainable.

Challenges, Risks, and Professional Practice

Common failures include confusing parameter distance with spatial distance, assuming control points are interpolated, using uniform tessellation for highly curved geometry, smoothing across creases, and simplifying without preserving boundaries. Numerical problems arise from nearly coincident points, repeated knots, degenerate faces, poorly conditioned parameterizations, and accumulated floating-point error.

A professional workflow should preserve source data, record algorithm parameters, use deterministic processing where possible, and validate every stage. Topological checks should identify nonmanifold vertices and edges, holes, disconnected components, and inconsistent orientation. Geometric checks should measure deviation, curvature changes, self-intersections, and element quality. Visual checks remain necessary because some artifacts are perceptually important but difficult to summarize with one metric.

There are also governance concerns. In industrial or medical applications, geometry may support safety-critical decisions, so provenance, versioning, tolerance documentation, and review procedures are essential. Scanned objects may contain sensitive environments or personally identifiable information. Assets used in virtual production or product design should respect licensing, ownership, and data-retention requirements.

Current Directions

Modern graphics pipelines increasingly combine analytic representations with GPU tessellation, displacement mapping, and multiresolution meshes. Patch-based rendering can delay polygon generation until visibility and screen size are known. Neural and differentiable representations are being investigated for reconstruction, inverse rendering, compression, and shape optimization, but they do not eliminate the need for topology, error bounds, interpretability, and robust export.

AI-assisted modeling can propose surfaces, repair holes, classify features, or predict simplification priorities. Its outputs should be treated as proposed geometry and evaluated against explicit tolerances. The most reliable systems combine learned priors with geometric constraints, traditional validation, and human review.

Best Practices

1. Choose a representation according to the required continuity, editability, accuracy, and downstream use.

2. Use local-support representations for large editable models and rational forms when exact conics are required.

3. Separate modeling tolerance from rendering tolerance.

4. Use adaptive tessellation and preserve shared patch boundaries.

5. Mark creases, boundaries, material seams, and semantic features before smoothing or simplification.

6. Maintain a high-resolution reference and generate controlled LODs from it.

7. Evaluate topology, geometry, parameterization, shading, and application performance together.

8. Record processing parameters so results can be reproduced and audited.

Chapter Summary

Parametric curves provide compact, controllable descriptions of shape. Bézier curves offer convex-hull behavior and straightforward subdivision; Hermite curves interpolate endpoints and tangents; B-splines provide local control through knots; and NURBS add weights and exact conic representation. Tensor-product, swept, lofted, and subdivision constructions extend these ideas to surfaces.

Continuous geometry must often be tessellated into meshes. Effective tessellation is adaptive and respects geometric error, screen-space importance, seams, and topology. Mesh processing then addresses parameterization, smoothing, simplification, deformation, and quality assessment. Each operation involves trade-offs between accuracy, smoothness, locality, performance, and feature preservation. Professional pipelines use explicit tolerances, multiresolution representations, automated validation, and application-specific decisions.

Key Terms

Parametric curve: A curve represented as a mapping from a parameter interval to spatial coordinates.

Control point: A point that influences, but does not necessarily lie on, a geometric curve or surface.

Bézier curve: A polynomial parametric curve expressed using Bernstein basis functions.

Bernstein polynomial: The basis function used to construct a Bézier curve.

De Casteljau algorithm: A recursive linear-interpolation method for evaluating and subdividing Bézier curves.

Hermite curve: A curve defined by endpoint positions and endpoint tangent vectors.

B-spline: A piecewise polynomial curve controlled by basis functions and a knot vector.

Knot vector: A nondecreasing sequence that determines parameter intervals and basis-function support.

NURBS: A Non-Uniform Rational B-Spline representation using knots, degrees, control points, and weights.

Continuity: The degree to which adjoining curve or surface pieces match in position, tangent, or curvature.

Tensor-product surface: A surface formed by combining two one-dimensional basis-function systems.

Tessellation: The conversion of continuous geometry into discrete line or polygon primitives.

Parameterization: The assignment of parameter-domain coordinates to geometric points.

Laplacian smoothing: A neighborhood-averaging operation used to reduce mesh variation.

Mesh simplification: Reduction of mesh complexity while attempting to preserve important properties.

Quadric error metric: An error measure based on distances from candidate geometry to associated face planes.

Level of detail: A representation selected according to viewing conditions or computational requirements.

Subdivision surface: A smooth limit surface generated by repeatedly refining a control mesh.

Nonmanifold geometry: Geometry whose local adjacency does not form a valid surface neighborhood.

Review Questions

Short Answer Questions

1. What property of Bernstein basis functions ensures that a Bézier curve lies within the convex hull of its control points?

2. State the geometric data required to define a cubic Hermite curve.

3. Why do B-splines provide more local control than a single high-degree Bézier curve?

4. What is the purpose of a knot vector in a B-spline?

5. Give two reasons why uniform tessellation may be inefficient.

Long Answer Questions

1. Derive and explain the endpoint tangent properties of a cubic Bézier curve.

2. Compare Bézier, Hermite, B-spline, and NURBS curves in terms of input data, continuity, locality, and applications.

3. Explain how tensor-product surfaces are constructed and discuss their regularity requirements.

4. Describe an adaptive surface-tessellation pipeline, including error criteria and crack prevention.

5. Explain mesh smoothing and simplification methods, including their principal artifacts and safeguards.

Analytical / Application Questions

1. A curved product shell contains flat panels, tight fillets, and marked crease lines. Design a tessellation and smoothing strategy that preserves appearance and geometric tolerances.

2. A character mesh must be reduced for mobile rendering while retaining facial features and joint deformation quality. Propose a simplification and LOD strategy and justify its constraints.

3. A scanned surface has uneven sampling, small holes, and nonmanifold regions. Develop a processing and validation workflow suitable for both measurement and real-time visualization.

22. Multiple Choice Questions with Answers

Question 1

Which property of Bernstein basis functions directly explains why a Bézier curve lies inside the convex hull of its control points?

A. They are orthogonal and have unlimited support.

B. They are nonnegative and sum to one.

C. They interpolate every interior control point.

D. They always produce uniform parameter speed.

Correct Answer: B. They are nonnegative and sum to one.

Question 2

A cubic Hermite segment is most directly specified by which data?

A. Four knots and four weights

B. Four control points only

C. Two endpoints and two endpoint tangent vectors

D. A polygon mesh and face normals

Correct Answer: C. Two endpoints and two endpoint tangent vectors

Question 3

What is the principal practical advantage of B-spline local support?

A. Every control point interpolates the curve.

B. Changing one control point affects only a limited region.

C. The curve always has constant curvature.

D. No knot vector is required.

Correct Answer: B. Changing one control point affects only a limited region.

Question 4

Why are NURBS especially useful in CAD systems?

A. They cannot represent polynomial segments.

B. They guarantee uniform triangle quality.

C. They can represent conic sections such as circles exactly.

D. They eliminate the need for tessellation in rendering.

Correct Answer: C. They can represent conic sections such as circles exactly.

Question 5

A renderer uses equal parameter intervals on a surface containing a flat area and a tight fillet. What is the most likely result?

A. Too many samples on the fillet and too few on the flat area.

B. Uniform geometric error everywhere.

C. Oversampling of the flat area and undersampling of the fillet.

D. Automatic preservation of all sharp features.

Correct Answer: C. Oversampling of the flat area and undersampling of the fillet.

Question 6

Which operation is most likely to shrink a mesh if applied repeatedly without correction?

A. Basic Laplacian smoothing

B. Knot insertion

C. De Casteljau subdivision

D. Quadric error evaluation

Correct Answer: A. Basic Laplacian smoothing

Question 7

In mesh simplification, a quadric error metric primarily estimates what?

A. Texture resolution at each vertex

B. Distance-related deviation from planes of affected faces

C. The number of connected components

D. The exact physical stress in the mesh

Correct Answer: B. Distance-related deviation from planes of affected faces

Question 8

Why should neighboring surface patches share consistent tessellation boundaries?

A. To guarantee constant curvature across the entire model

B. To prevent cracks and T-junctions between independently refined patches

C. To make all triangles equilateral

D. To remove the need for surface normals

Correct Answer: B. To prevent cracks and T-junctions between independently refined patches

Question 9

Which parameterization is generally more related to actual spacing than uniform parameterization?

A. Chord-length parameterization

B. Random parameterization

C. Degree-zero parameterization

D. Weight-free parameterization

Correct Answer: A. Chord-length parameterization

Question 10

A mesh-processing pipeline must preserve a machined crease while removing scan noise. Which strategy is most appropriate?

A. Apply unconstrained global smoothing across all edges.

B. Delete all high-curvature regions.

C. Mark the crease as a feature and use feature-aware smoothing.

D. Use uniform decimation without boundary penalties.

Correct Answer: C. Mark the crease as a feature and use feature-aware smoothing.