Computer Graphics Forum
  • Preprint
  • Article type: Major Revision from Eurographics Conference
  • Funding information: National Science Foundation, IIS‐2402893; Natural Sciences and Engineering Research Council of Canada, RGPIN‐2024‐03981

SharpEdge: Edge Feature Surface Reconstruction From Space Curves

Abstract

Reconstructing high‐quality surfaces from sparse 3D space curves remains challenging, particularly when sharp feature edges are involved. Existing methods often reconstruct smooth surfaces that ignore sharp edges, depend on a failure‐prone automatic initialization procedure, or rely on manual intervention for initial surface estimation, all of which can lead to a loss of intended structure. We introduce SharpEdge, a framework for edge‐aware surface reconstruction from sparse (possibly disconnected) curve networks. Our method optimizes a two‐normal field along curve edges and enforces geometric constraints at curve junctions to better capture sharp feature directions. Guided by these normals, we reconstruct an initial surface using an adapted signed‐distance heat method [FC24] and refine it with a cut–remesh–optimize pipeline that produces fair surface patches interpolating the features. In our experiments on a range of 3D sketches, SharpEdge produces meshes that adhere closely to input curves and recover sharp ridges and corners that are often smoothed out by alternative approaches.
Keywords: curve networks, feature curves, geometry processing, shape modelling, surface reconstruction

1 Introduction

Curves in three‐dimensional space are a natural and expressive medium for describing shape. They arise in many design workflows, including CAD modelling, industrial design sketches, and immersive VR drawing [YAS*21, YDSG21, RRS19]. Compared to dense point clouds or meshes, curve networks are compact and intuitive, but their sparsity and irregular connectivity make surface reconstruction especially difficult. In particular, many applications expect sharp edges—from functional boundaries in CAD models to structural lines in sketch‐based design—yet existing reconstruction methods often oversmooth these features.

1.1 Point‐Based Reconstruction

A classic and still widely used approach to point‐based surface reconstruction is the Poisson surface reconstruction algorithm [KBH06]. More recently, approaches have been proposed based on generalized winding numbers [BDS*18] and the heat method [FC24]. However, these approaches requires an oriented normal for each point sample. While the curve input that we study can be trivially sampled to obtain points, there is no clear method for obtaining surface normals.
There has been a flurry of recent approaches for surface reconstruction from point clouds without normals [HCJ19, HWW*22, XDW*23, LLC*25, XJ25, SCC*25, CB24]. In contrast with our approach, none of these works consider sharp features, leading to oversmoothed output. Moreover, curve input exhibits extreme sparsity away from the curves (e.g., the empty faces of a box), which can lead these approaches to create surfaces that balloon outward or form undesirable tunnels or other topological artifacts. See, for example, our Figure 15, [XJ25] Figure 12, [WDL*25] Figure 11, [SPB*25] Figure 14, and [XDW*23] Figure 18.
Figure 1
(a) Input 3D curve network. (b) Optimization of a two‐normal field along curve edges. (c) Initial surface reconstruction guided by the estimated normals using the generalized signed‐heat method. (d) Final surface after a cut–remesh–optimize pipeline producing well‐shaped patches that interpolate the features.

1.2 Curve‐Based Methods

An alternative is to reconstruct surfaces directly from curve networks, leveraging their connectivity and geometry. Classical approaches identify cycles in curve networks [ZZCJ13, AJA11] and triangulate them into surface patches [ZJC13, AJG*13]. These approaches do not produce smooth surfaces. Flow‐aligned surfacing [PLS*15] classified strokes into flow lines and trim curves to guide patch construction, producing high‐quality outputs when the input is well‐structured (a fully connected curve network with clean topology). Contemporaneously to our work, Variational Neural Surfacing [ZWB25] builds on this idea by reconstructing implicit neural surfaces from clean or rough 3D sketches using curvature‐aligned flow‐line priors inspired by perceptual observations. SurfaceBrush [RRS19] creates manifold surfaces from freehand VR ribbons by exploiting local tangent coherence, but assumes dense, surface‐covering input. Yu et al. [YAB*22] proposed the first method to handle unstructured 3D sketches. It depends on a proxy surface for initialization. They proposed to use VIPSS [HCJ19] to automatically generate the proxy, but often resort to manual proxy creation. More recently, VRSurf [SPB*25] inflates sparse strokes into closed surfaces, but produces smooth results that miss intended sharp features (Figure 16).

1.3 Data‐Driven Methods

Data‐driven approaches have also been proposed to create neural implicit surfaces from unoriented point clouds [HNSS*25, FSYL23, WZX*23, EFPH*24, EGO*20]. These approaches rely on large training datasets and come with heavy runtime costs. Although some approaches are capable of reconstructing sharp features [EFPH*24, EGO*20, FSYL23], they all make the same uniform input density assumption as point‐based reconstruction approaches. As such, they are not capable of handling the sparsity of curve‐based input, as shown in [WDL*25] Figure 6. One concurrent work, NeuVAS [WDL*25] is designed for curve‐based input and produces high‐quality results. However, NeuVAS requires the user to manually specify sharp features.

1.4 Our Approach

We introduce SharpEdge, a framework for edge‐aware surface reconstruction from sparse (possibly disconnected) curve networks. Our key idea is to estimate a two‐normal field along curve edges. We do this by optimizing the normal smoothness independently on both sides of curves, with geometric constraints and a relaxed notion of incidence at curve junctions to help preserve sharp features. Guided by these normals, we reconstruct an initial surface using the signed‐heat distance method [FC24] which we adapt to two‐normal curve input, and then apply a cut–remesh–optimize pipeline to improve mesh quality while maintaining feature boundaries.
Our contributions are:
•
We propose a two‐normal field representation for curve edges and an accompanying notion of incidence at junctions, allowing explicit modelling of sharp features that cannot be captured by a single normal.
•
We adapt the signed‐heat distance method to curve‐guided surface reconstruction with two normals, enabling implicit surfacing from sparse sketch input.
•
We design a cut–remesh–optimize pipeline that improves mesh quality while preserving feature boundaries, producing surfaces suitable for downstream geometry processing.

2 Overview

The input to SharpEdgeis a curve network represented as a 3D polyline graph. The data consists of:
•
A set of 3D vertices V={v1,v2,⋯,vn}⊂R3$V = \lbrace \mathbf {v}_1, \mathbf {v}_2, \dots, \mathbf {v}_n \rbrace \subset \mathbb {R}^3$.
•
A set of polylines L={L1,L2,⋯,LK}$\mathcal {L} = \lbrace L_1, L_2, \dots, L_K \rbrace$, where each Li$L_i$ is an ordered sequence of vertex indices (open or closed).
Different polylines may be connected via shared vertices, and the overall network may contain sharp junctions or T‐junctions and even be disconnected. We assume that the curves lie on an underlying 2‐manifold surface M$\mathcal {M}$, possibly with sharp features. Our goal is to reconstruct a surface that approximates M$\mathcal {M}$ and adheres to the input curves, in particular preserving the implied sharp features.
To achieve this goal, our algorithm proceeds in several stages. First, we estimate candidate normals for some curve edges from the convex hull and extend them along curves via parallel transport (Figure 2, §3). These normals are refined through a global single‐normal optimization (§4), followed by constraint extraction at high‐valence vertices and a subsequent two‐normal optimization that computes dual normals per edge (§5). Guided by this dual‐normal field, we reconstruct an implicit surface using an adapted signed heat method [FC24] and extract a mesh. Finally, we cut the mesh into patches along curve edges, remesh, and perform per‐patch normal‐guided smoothing [SHBSS16] to improve mesh quality and faithfully preserve sharp features along the input curves (§6).
Figure 2
Parallel transporting the normal plane along a curve provides a frame which undergoes minimal rotation [HM95, Bis75]. At any point along a curve, its unique tangent T$T$ and any vector N$N$ normal to it defines a local frame. (In 3D, a pair of orthogonal vectors in the normal plane are needed.) We transport the frame via the minimal rotation which aligns tangent vectors. This keeps normal vectors perpendicular to the tangent. In this way, any vector v$v$ in a local frame can be transported along the curve.

3 Initial Normal Estimation

Given a collection of space curves that lie on an implied manifold, our goal is to find an estimate of the surface normal along the curves. These estimated normals will serve as an initial guess and regularization for subsequent non‐linear optimization. The most important property of the estimated normals is that they should be outward facing, since flipping all normals would also lead to a smooth (but incorrect) solution. Our key insight is that, for a manifold, curves on the convex hull must have exterior‐facing normals.
We begin by computing the convex hull of the input vertex set V$V$ (Figure 3a,b) (scaled to lie within the unit cube (§8.1).) We extract outward‐pointing normals for all convex hull faces. The convex hull's edges and vertices are assigned normals by averaging the normals of adjacent faces. These normals are used to assign candidate normals to edges in the input curve network (Figure 3c). Specifically, if an edge lies directly on the convex hull, we assign it the corresponding hull edge normal. If either of its endpoints lies on the convex hull, we use the hull vertex normal. If the edge lies within a small threshold (i.e., <5×10−3$< 5 \times 10^{-3}$) of any convex hull face, we assign an average of the within‐threshold face normals. Finally, we discard normals that are nearly parallel to the edge direction (within 5∘$5^\circ$ of parallel), as they are unreliable for surface construction. Let H={(ei,ni0)}$\mathcal {H} = \lbrace (e_i, n_i^0) \rbrace$ be the set of edges ei$e_i$ with valid normals ni$n_i$.
Figure 3
Visualization of our initial normal estimation pipeline. The resulting initial estimate assigns one normal to every edge, ready for downstream optimization. From left to right: (a) the input sketch consists of multiple coloured polylines; (b) the convex hull of the vertex set; (c) normals extracted from convex hull faces are assigned to nearby edges; (d) convex hull normals are propagated via parallel transport along each curve; (e) normals are iteratively propagated to and along nearby curves until every edge is reached.
For polylines that contain one or more valid edge normals (i.e., elements of H$\mathcal {H}$), we propagate one normal along each curve using parallel transport [HM95, BWR*08]. The propagation starts from the edge with the most perpendicular normal and proceeds along the polyline while preserving orthogonality (Figure 3d).
For polylines that still lack normals, we apply an iterative propagation strategy. Beginning with the endpoints of polylines that already have assigned normals (the ‘frontier'), we identify neighbouring edges without normals. The frontier normals are propagated to the adjacent edges and then along their respective polylines. This process is repeated until no further normals can be assigned (Figure 3e).

4 Single‐Normal Optimization

Our initially estimated normals are assumed to have the correct inside‐outside orientation, but may not minimize surface smoothness or meet smoothly at curve junctions. Our goal is to refine the normals so that they vary smoothly (imply a smooth surface) along the curve and at curve junctions. We do this via a global optimization that takes parallel transport into account. Following this optimization, each edge ei$e_i$ is assigned a single angle parameter θi$\theta _i$, which specifies its normal direction within an arbitrary local orthogonal frame (ui,vi)$(\mathbf {u}_i, \mathbf {v}_i)$. See Figure 4 for an overview.
Figure 4
One‐normal optimization. (a) Initial normals estimated from convex hull propagation. (b) Each edge normal is represented by an angle θi$\theta _i$ in its local frame (ui,vi)$(\mathbf {u}_i, \mathbf {v}_i)$. (c) Normals evolve during gradient‐based optimization to be smooth along connected edges. (d) The final configuration yields coherent, outward‐facing normals across the entire curve network.
We formulate an energy Eone-normal$E_{\text{one-normal}}$ over these angles with two terms,
Eone-normal(θ)=10−2Ereg(θ)+104∑i,j∈EEpairwise(θi,θj),$$\begin{align} E_{\text{one-normal}}(\theta) = 10^{-2} \, E_{\text{reg}}(\theta) + 10^4 \, \sum _{{i,j \in \mathcal {E}}} E_{\text{pairwise}}(\theta _i, \theta _j), \end{align}$$
(1)
where i,j∈E$i,j \in \mathcal {E}$ identifies incident edges.
The regularization term Ereg$E_{\text{reg}}$ encourages, for edges in H$\mathcal {H}$, its optimized angle θi$\theta _i$ to remain close to its initial estimated normal ni0$n_i^0$ from the convex hull estimation. This is measured as:
Ereg(θ)=∑ei,ni∈H(1−ni·ni0)2$$\begin{align} E_{\text{reg}}(\theta) = \sum _{e_i, n_i \in \mathcal {H}} (1 - n_i \cdot n_i^0)^2 \end{align}$$
(2)
where ni=cos(θi)ui+sin(θi)vi$n_i = \cos (\theta _i) \mathbf {u}_i + \sin (\theta _i) \mathbf {v}_i$.
The pairwise term Epairwise$E_{\text{pairwise}}$ promotes consistency between incident edges ei$e_i$ and ej$e_j$, penalizing angular deviation between edge normals. Directly comparing the associated angles would incorrectly penalize the geometric turning of the curve. Instead, we wish to penalize any twisting of the surface beyond the intrinsic curvature of the curve. In other words, we prefer zero local Gaussian curvature. To achieve this, we penalize deviation from the parallel transport rotation Δi→j$\Delta _{i \rightarrow j}$ when evaluating the pairwise term,
Epairwise(θi,θj)=(1−Δi→jni·nj)2,$$\begin{align} E_{\text{pairwise}}(\theta _i, \theta _j) = (1 - \Delta _{i \rightarrow j} n_i \cdot n_j)^2, \end{align}$$
(3)
where Δi→j$\Delta _{i \rightarrow j}$ is the minimal rotation that maps ei$e_i$ to ej$e_j$, and ni$n_i$ and nj$n_j$ are constructed from θi$\theta _i$ and θj$\theta _j$ in their respective local frames. Let a$\mathbf {a}$ be the parallel transport axis of rotation (i.e., the cross product of ei$e_i$ and ej$e_j$, which is also the plane normal for a planar curve). Since ni$n_i$ and nj$n_j$ are orthogonal to their edges and compared after parallel transport, the zero‐cost configurations correspond to normal directions that are consistent with both edges. Two (unoriented) normal directions have zero cost. One is parallel to a$\mathbf {a}$. The other is the normal for a cylinder with axis a$\mathbf {a}$ passing through the three vertices ei$e_i$ and ej$e_j$.

5 Two‐Normal Optimization

In our one‐normal optimization (§4), we computed a smoothly varying normal ni$n_i$ for each curve edge ei$e_i$. This normal was based on the assumption that the curves lie on a smooth manifold surface. Our goal in this stage is to relax the smoothness assumption and allow for the surface to crease along curve edges. We do this by optimizing a generalized version of our Eone-normal$E_{\text{one-normal}}$ objective function that allows for two (possibly distinct) normals per edge (niA$n_i^A$ and niB$n_i^B$). First, we define smoothness along a curve such that the implied surface along each side is smooth yet can be sharp across the edge, without knowing a priori whether to compare the A$A$ or B$B$ normals. Second, we modify our definition of edge incidence to allow for sharp features at valence >2$>2$ junctions. Third, we observe that piece‐wise smooth corners of the surface meet at valence >2$>2$ junctions and imply well‐defined normal constraints.

5.1 Normal Matching

Each edge carries two candidate normals. To define an Esmoothness$E_{\text{smoothness}}$, we must define a notion of smoothness between incident edges that promotes smoothness along each side of the curve while allowing for a normal discontinuity across the edge. At valence‐2 vertices, the two incident edges lie on a continuous curve segment. We desire a smooth surface on both sides of the edges, without assuming an orientation. We evaluate the smoothness of the better match among the two possible matches (Figure 5, left). Namely, for a pair of edges ei,ej$e_i,e_j$ that meet at a valence‐2 junction:
Esmoothness(ei,ej)=minEpairwiseθiA,θjA+EpairwiseθiB,θjB,EpairwiseθiA,θjB+EpairwiseθiB,θjA$$\begin{align*} E_{\text{smoothness}}(e_i,e_j) = \min {\left(E_{\text{pairwise}}{\left(\theta _i^A,\theta _j^A\right)} + E_{\text{pairwise}} {\left(\theta _i^B,\theta _j^B\right)}, \right.} \\ {\left. E_{\text{pairwise}}{\left(\theta _i^A,\theta _j^B\right)} + E_{\text{pairwise}}{\left(\theta _i^B,\theta _j^A\right)} \right)} \end{align*}$$
Figure 5
Normal assignment at valence‐2 and higher vertices. (Left) At a valence‐2 vertex, each edge has two candidate normals (nA,nB$n^A,n^B$). Our optimization dynamically selects the more consistent of the two possible matches, {n0A↔n1A,n0B↔n1B}$\lbrace n_0^A \leftrightarrow n_1^A,n_0^B \leftrightarrow n_1^B \rbrace$ or {n0A↔n1B,n0B↔n1A}$\lbrace n_0^A \leftrightarrow n_1^B,n_0^B \leftrightarrow n_1^A \rbrace$, yielding smooth normals along either side of the curve. (Right) At a higher‐valence junction, evaluating all pairs of incident edges would oversmooth a potentially sharp corner. Instead, we compute a circulation and a consistent comparison strategy that only compares a single normal between circulation‐adjacent edges (niB↔ni+1A$n_i^B \leftrightarrow n_{i+1}^A$). The two normals of each edge are used consistently for each incident corner of the surface, ensuring local smoothness while preserving sharp discontinuities.
At a valence >3$>3$ vertex, we interpret all incident curves as lying on the surface (Figure 6a). However, since we only assume our surface to be piecewise‐smooth, we don't wish to compare edges across a sharp feature. Therefore, we establish a local circulation around the vertex by sorting the edges counter‐clockwise (§5.2 and Figure 6b). Each circulation‐adjacent pair of edges bounds a smooth corner of the surface. For a pair of circulation‐adjacent edges ei,ej$e_i,e_j$ that meet at a valence >2$>2$ junction, we only compare a single normal between them, as determined by the circulation order (Figure 5, right):
Esmoothness(ei,ej)=Epairwise(θiB,θjA).$$\begin{align} E_{\text{smoothness}}(e_i,e_j) = E_{\text{pairwise}}(\theta _i^B,\theta _j^A). \end{align}$$
(4)
Figure 6
Edge circulation and constraint construction at a high‐valence vertex. Here, “circulation” denotes the cyclic ordering of incident edges in the local tangent plane. (a) Four incident edges (18, 33, 42, 79) are shown. (b) We order the edges into a directed circulation. Each circulation‐adjacent pair of edges defines a smooth corner. The circulation order determines smoothing pairs: (e18,e33)$(\mathbf {e}_{18}, \mathbf {e}_{33})$, (e33,e42)$(\mathbf {e}_{33}, \mathbf {e}_{42})$, (e42,e79)$(\mathbf {e}_{42}, \mathbf {e}_{79})$, and (e79,e18)$(\mathbf {e}_{79}, \mathbf {e}_{18})$. (c) The two normals for each edge are constrained based on cross products with its circulation‐adjacent edges. For example, edge 18 yields e18×e33$\mathbf {e}_{18} \times \mathbf {e}_{33}$ and e79×e18$\mathbf {e}_{79} \times \mathbf {e}_{18}$.
Since circulation‐adjacent edges bound smooth corners, their cross product provides a known surface normal. Therefore, we also define a constraint term:
Econstraint(ei,ej)=(1−niB·cij)2+(1−njA·cij)2,$$\begin{align} E_{\text{constraint}}(e_i,e_j) = (1 - n_i^B \cdot \mathbf {c}_{ij})^2 + (1 - n_j^A \cdot \mathbf {c}_{ij})^2, \end{align}$$
(5)
where cij$\mathbf {c}_ij$ is the unit‐length vector parallel to ei×ej$e_i \times e_j$. To resolve the sign ambiguity, we leverage the one‐normal optimization and orient cij$\mathbf {c}_ij$ to be outward‐facing according to ni+nj$n_i+n_j$, the edge normals obtained from minimizing Eone-normal$E_{\text{one-normal}}$. Nearly parallel edge pairs (i.e., T‐junctions) are skipped to ensure stability (threshold angle 10∘$10^\circ$). See Figure 6 for an example. To avoid conflicting constraints from both ends of an edge, we split at the midpoint any edge with both endpoints’ valences >2$>2$.

5.2 Edge Circulation for High‐Valence Vertices

An edge circulation at a vertex with valence >2$>2$ provides the local topological information needed to establish surface corners, and with them pairwise smoothness and normal constraints. To compute the circulation, we find a tangent plane passing through the vertex and sort projected edge vectors by angle. We estimate the tangent plane normal n$\mathbf {n}$ by averaging the optimized incident edge one‐normals at the vertex.1

5.3 Optimization

Our two‐normal optimization energy is defined as:
Etwo-normal(θA,θB)=∑i,j∈E′Esmoothness(ei,ej)+∑Econstraint(ei,ej),$$\begin{align*} E_{\text{two-normal}}(\theta ^A, \theta ^B) = \sum _{{i,j \in \mathcal {E}^{\prime }}} E_{\text{smoothness}}(e_i, e_j) + \sum E_{\text{constraint}}(e_i, e_j), \end{align*}$$
where i,j∈E′$i,j \in \mathcal {E}^{\prime }$ is the set of valence‐2‐incident edges and circulation‐adjacent edges, and ∑Econstraint(ei,ej)$\sum E_{\text{constraint}}(e_i, e_j)$ sums over circulation‐adjacent edges. This expression incorporates high‐valence corner smoothness and normal constraints, with orientations resolved using the one‐normal optimization result. It measures smoothness along either side of edges, while allowing for sharp edges and taking care not to oversmooth high‐valence junctions. The initial guess for θA$\theta ^A$ and θB$\theta ^B$ at edges incident to valence >2$>2$ vertices is provided by the corner normal. The initial guess elsewhere is the one‐normal solution, perturbed. Figure 7 illustrates this optimization.
Figure 7
Workflow and results of two‐normal optimization. (a) Initialization with one normal perturbed from the one‐normal solution. (b) High‐valence vertex constraints guide the optimization. (c–d) The normal field evolves during optimization and converges after 34 iterations. (e–f) The final result is separated into two consistent normal groups by similarity‐based clustering.

6 Initial Surfacing

Once a two‐normal field has been computed, the next step is to recover a surface consistent with these normals. We decompose this task into two sub‐problems. First, we construct a smooth, globally defined signed distance function whose zero level set interpolates our curve input with a topology matching our desired output. We then optimize the resulting surface to smoothly interpolate the edge normals.

6.1 Implicit Surface Computation

We require a method that is robust to sparse input, capable of interpolating constraints, and can be adapted to our two‐normal scenario. For this reason, we adopt the signed heat method (SHM) of Feng and Crane [FC24]. This method diffuses orientation information across a volumetric discretization to obtain a consistent signed distance field.
In contrast, Poisson Surface Reconstruction (PSR) assumes dense oriented point samples and produces smooth approximations that do not strictly interpolate sparse curve constraints. Neural SDF‐based approaches, while flexible, typically rely on dense sampling and optimization, do not guarantee exact interpolation, and are not readily adaptable to our two‐normal setting.
The original SHM formulation only supports a single normal per sample and is designed for point or surface inputs. We extend it to handle two normals per sample. (We also experimented with generalized winding number approaches [BDS*18], but did not find them as adaptable to our two normals scenario.)
In the signed heat method, normals are first propagated via short‐time heat diffusion. Intuitively, this step spreads the input normals throughout the volume to form a smooth and coherent orientation field. Specifically, a diffused vector field X$\mathbf {X}$ is obtained by integrating element normals with the Yukawa potential (and renormalizing to unit length). The method then solves a Poisson equation throughout the domain to find a function whose gradient best matches X$\mathbf {X}$. The resulting implicit function defines a signed distance field.
We adapt SHM to curve networks with dual‐normal constraints in two ways. First, we treat each curve edge as a linear source element that contributes orientation constraints to the diffusion process. Second, we modify edge elements to propagate two normals. At query time, we select between them using a geometric criterion based on the relative position of the query direction with respect to the normal “cones.” These extensions allow the signed‐heat formulation to incorporate curve‐based dual‐normal inputs and faithfully propagate sign constraints that preserve sharp features.

6.1.1 Edge Adaptation

To apply SHM to curve‐based inputs, we must reinterpret orientation constraints so that they are imposed along 1D curve edges rather than 0D points or 2D surface elements. In the original SHM, orientation constraints are integrated over points or surface elements with associated normals. For our edge‐based dual normal geometry, we instead treat each curve edge as a linear source element. Each edge ei$e_i$ connecting vertices pi$\mathbf {p}_i$ and pi+1$\mathbf {p}_{i+1}$ contributes to the diffused vector field through line integration along its length, ensuring that curve edges directly constrain the heat diffusion process.

6.1.2 Dual‐Normal Adaptation

The dual‐normal adaptation ensures that the two distinct normal directions associated with each curve edge are properly respected during diffusion. When integrating element normals, we need to determine which of the two normals n1,n2$\mathbf {n}_1,\mathbf {n}_2$ to diffuse to a query point q$\mathbf {q}$. We interpret the edge element and its normals as defining a sharp, convex surface (Figure 8), and diffuse the normal of its closest point. Formally, let p$\mathbf {p}$ be the edge midpoint. Let d$\mathbf {d}$ be the query direction q−p$\mathbf {q} - \mathbf {p}$ projected onto the plane orthogonal to the edge. We apply a criterion based on the query direction d$\mathbf {d}$ relative to the half‐spaces defined by the normals:
•
If d$\mathbf {d}$ is outside one half‐space but inside the other, the query point is closest to the outside half‐space (i.e., d·n1>0,d·n2<0$\mathbf {d}\cdot \mathbf {n}_1 > 0,\ \mathbf {d}\cdot \mathbf {n}_2 < 0$ or vice versa). We take the normal of the outside half‐space.
•
If d$\mathbf {d}$ lies outside both half‐spaces (i.e., d·n1>0,d·n2>0$\mathbf {d}\cdot \mathbf {n}_1 > 0,\ \mathbf {d}\cdot \mathbf {n}_2 > 0$), we compare against the bisector n1+n2∥n1+n2∥$\tfrac{\mathbf {n}_1+\mathbf {n}_2}{\Vert \mathbf {n}_1+\mathbf {n}_2\Vert }$ to determine if it lies within the wedge spanned by n1$\mathbf {n}_1$ and n2$\mathbf {n}_2$. If d$\mathbf {d}$ is closer to the bisector, then we the use the radial direction d∥d∥$\tfrac{\mathbf {d}}{\Vert \mathbf {d}\Vert }$ as the normal; otherwise, we choose whichever normal is more aligned with d$\mathbf {d}$.
•
Otherwise (both dot products negative), the query point lies within both half‐spaces. We pick the normal that is more aligned with d$\mathbf {d}$, which chooses the normal of the closest point on either half‐space.
Figure 8
Illustration of our SHM normal diffusion criterion. Two normals n1,n2$\mathbf {n}_1,\mathbf {n}_2$ span a cone at edge midpoint pi$\mathbf {p}_i$ (viewed along the edge). The bisector b$\mathbf {b}$ and query q$\mathbf {q}$ are shown, and the shaded grey region suggests the implied sharp surface. The selected normal is chosen to best capture this sharpness.
This criterion exactly matches the level‐set behaviour of signed distance fields near convex sharp features. It may, however, round sharp concave features, which is a direction for future work (the iron in Figure 20).
We run SHM using a constrained Delaunay tetrahedralization computed via tetgen [Si15], which explicitly embeds all curve edges to preserve them in the output. We then extract the zero level set using marching tetrahedra.
During isosurface extraction, tetrahedral edges may occasionally collapse or disconnect, causing topological inconsistencies or artifacts. To improve robustness, we explored several remedies, including (i) adjusting the tetrahedral resolution; (ii) fitting a smooth cubic approximation of the scalar field along heterogeneous edges, and inserting additional sample points along edges where the cubic fit detects an undiscretised sign flip (re‐running tetgen and SHM); and (iii) perturbing function values during the isovalue sign test only to alter topology while preserving geometry.
In practice, we settled on a solution that works for most examples in our dataset, though it does not eliminate all failures. We constrained TetGen with a minimum tetrahedron volume proportional to the cube of the mean curve edge length. This worked for all but three models that we tried. If this setting fails to capture all curve edges, we first resample the edges at a different resolution (spherecylinder, bowl_curve); if this also fails, we apply a slight perturbation to the input. Among all models in this paper, only the toothpaste example (Figure 2) required perturbation.
Although these strategies alleviate many issues, marching tetrahedra remains the least robust component of our pipeline (Figure 9).
Figure 9
Failure cases of marching tetrahedra. Mouse: an erroneous dimple; despite this, cut–remesh–optimize can still recover a nice surface. Vacuum: the initial surface is missing a curve edge. Toothpaste: an ear triangle becomes non‐manifold and blocks remeshing. Even when the result is not manifold or even missing edges, cutting and optimizing will improve the surface quality, though element quality remains poor (Section 8.7). While adjusting input resolution or perturbations can sometimes alleviate these issues, they may still arise in several cases.

7 Surface Optimization

The surface extracted from marching tetrahedra typically suffers from two issues. First, it often contains degenerate edges and triangles, where repeated or nearly coincident vertices make curve edges difficult to identify. Second, it may contain artifacts such as dents and bumps and does not minimize an objective fairness criteria. To alleviate these issues and obtain a piece‐wise fair surface, we remesh the surface and optimize positions to have piecewise‐fair normals.
To ensure that curve edges can be reliably identified for subsequent processing, we begin by collapsing zero‐length edges. Next, to allow for further piecewise smooth processing, we cut the surface along the input curve edges. We then perform isotropic remeshing using CGAL's isotropic surface remeshing routine [BKP*10], which redistributes vertices, equalizes edge lengths, and improves triangle quality while respecting boundaries. As we did not implement a link condition test while collapsing zero‐length edges, the isotropic remeshing step sometimes fails. However, our surface optimization algorithm is robust to mesh element quality and still succeeds.
Each surface patch is then optimized using the normal‐driven fairing method of Stanko et al. [SHBSS16]. Naive bi‐Laplacian smoothing using a linearized Laplace operator produces pinched cylinders unlike a non‐linear bending energy. To avoid these issues, Stanko et al. [SHBSS16] couple two curvature identities [Sul07]:
Δx=2Hnand∫∫DHndA=∮∂D(n×e)ds,$$\begin{align} \Delta \mathbf {x} = 2 H \mathbf {n} \quad \text{and} \quad \iint\nolimits _D H\mathbf {n} \, dA = \oint _{\partial D} (\mathbf {n} \times \mathbf {e}) \, ds, \end{align}$$
(6)
where n$\mathbf {n}$ is the surface normal and e$\mathbf {e}$ the unit tangent vector along the boundary. The optimization is realized as three sparse linear solves: (i) propagating boundary positions subject to boundary normal constraints to create a smooth, initial parametric domain; (ii) propagating a smooth normal field subject to boundary normal conditions, and (iii) updating vertex positions to match the smoothed normals via Equation 6. For robustness, we employ the mollified intrinsic Laplace operator [SC20], which provides a stable discretization even on irregular meshes. To stabilize the operator near our open cut boundaries and provide normal boundary constraints to step (i), we augment the mesh with equilateral flap triangles aligned with the boundary normals (Figure 10). (Each flap triangle is attached to a boundary edge and acts as a local auxiliary face that preserves the boundary normal direction during optimization. Flap triangles are deleted after optimization.) Each cut edge takes the normal most aligned with its incident face. In our two‐normal setting, each curve edge is associated with a pair of normals. After cutting the mesh along these edges, each side of the cut naturally supports its own boundary conditions. For smooth edges, the two normals coincide after optimization, and identical constraints are applied to both sides, ensuring continuity. For sharp edges, the two normals differ, and distinct constraints are enforced on either side of the cut, preserving the intended crease geometry.
Figure 10
Optimization with flaps. (a) Input surface with boundary flaps (red vertices and equilateral triangles). (b) Smooth propagation of normals across the patch under boundary constraints. (c) Final optimized surface, where curvature‐based adjustment yields a smooth interior while flaps preserve sharp curve boundaries.
This approach integrates seamlessly with the Stanko et al. [SHBSS16] approach, allowing us to obtain piecewise smooth surface patches that interpolate the curves and the normal on each side, without blurring across sharp features. Our cut–remesh–optimize pipeline (Figure 11) yields meshes that are both feature‐preserving and well‐suited for downstream geometric processing.
Figure 11
Cut–remesh–optimize pipeline for the spherecylinder model. (a) The mesh is cut along the input curve edges. (b) The cut surface is isotropically remeshed to improve element quality while respecting boundaries. (c) Each surface patch is optimized using intrinsic smoothing, removing noise without blurring across sharp edges. The resulting mesh is fair and feature‐preserving with well‐shaped elements.

8 Results

8.1 Preprocessing

We evaluated our approach on curve networks from [XCS*14, PLS*15, GSV*17, YDSG21], which we preprocess through the following steps. We first remove duplicate vertices and curves, and split curves at intersections to obtain connected polylines. The polylines are then simplified using the Ramer–Douglas–Peucker algorithm [DP73] to reduce the vertex count while preserving the overall shape. Finally, each sketch is centred and uniformly scaled to unit size.

8.2 Implementation

Our normal optimization (§4,§5) is implemented in Python with SciPy and JAX. We minimize Eone-normal$E_{\text{one-normal}}$ and Etwo-normal$E_{\text{two-normal}}$ using the L‐BFGS‐B method with automatic differentiation. The surface optimization (§7) is written in Python with SciPy and libigl [JP*18]. It is formulated as a sparse linear system and solved using UMFPACK. Cutting and remeshing operations are implemented in C++ with CGAL and Eigen. Our modified signed heat method is based on the reference implementation of Feng et al. [FC24].
All experiments were conducted on a MacBook Pro (14‐inch, 2021) with an Apple M1 Pro chip (10 cores: 8 performance and 2 efficiency) and 16 GB of RAM.

8.3 Running Time

On the sewing machine example (Figure 1), which contains 392 curve edges and 360 vertices, normal optimization (§3–5) takes about 10 s, the single‐normal optimization runs for 204 BFGS iterations, and the dual‐normal optimization converges in 44 iterations. Initial surface generation (§6) about 14 s, and surface optimization (§7) about 5 s. For more complex inputs, the times may increase but remain below 30 s for each stage. Our overall pipeline typically finishes within one minute.

8.4 Examples

Figures 1, 15, 16, and 21 show a gallery of surfaces created using SharpEdge. An interactive gallery is provided in the supplemental materials. The reconstructions adhere to the sketches while producing watertight, well‐shaped meshes that capture sharp edge features. The two normals naturally and automatically unify into a single normal or become bimodal and follow different incident surfaces as appropriate.
Figure 16
Comparison between VRSurf [SPB*25] and our method. VRSurf produces an overly smooth surface, whereas our approach automatically finds and preserves sharp edges. Our fair surface optimization recovers cylinders and spheres.

8.5 Comparisons

We compare our method against several prior approaches, including VIPSS [HCJ19], Yu et al. [YAB*22], and the recent VRSurf [SPB*25]; see Figs. 15,16.
Using a single normal field implicitly assumes the absence of sharp features, which becomes especially evident for sparse curve networks. Yu et al. [YAB*22] reconstructs piecewise‐smooth surfaces starting from an initial proxy surface, and is able to enhance sharp features. However, this method is sensitive to the initialization: for example, in the toothpaste case (Figure 15), it requires a manually provided proxy. Moreover, since the output is only piecewise‐smooth, transitions between patches are C0$C^0$ continuous but lack tangent control. For example, the lack of tangent continuity (G1$G^1$) can be seen at the front junction of the mouse. Many sharp edges exhibit a lack of precise tangent control (e.g., the top junction of the toothpaste model).
Zhuang et al. [ZZCJ13] addressed the problem of cycle discovery in 3D curve networks, aiming to identify closed curve loops that may correspond to potential surface patches. Their method focuses exclusively on cycle extraction and does not perform surface reconstruction. Moreover, the formulation relies on a rotation graph to identify valid cycles, an assumption that becomes ill‐defined in certain configurations. For example, Figure 12 shows a configuration in which the rotation‐graph formulation fails. The method also does not provide a clear extension to disconnected or partially disconnected curve networks. Our work combines both cycle discovery and surfacing into a unified approach. We don't explicitly extract cycles or perform patch‐wise surface fitting. Instead, we directly compute a two‐normal field along the curves and then reconstruct a surface interpolating the curves and normals. Our approach works on disconnected components (discussion below).
Figure 12
An illustrative example in which explicit cycle extraction under the rotation‐graph formulation of [ZZCJ13] fails. (a) The input sketch consists of two small circular curves near the north and south poles of a sphere, connected by a single spiral curve. (b) The two‐normal field generated by our method. (c) Our optimized surface closely reconstructs a sphere.
Pan et al. [PLS*15] build upon the cycle detection of Zhuang et al. [ZZCJ13] by reconstructing surfaces from detected cycles, followed by smoothing, remeshing, and curvature‐direction alignment to flow‐line curves. As a result, their pipeline inherits the assumptions and limitations of explicit cycle extraction. In cases where cycle detection is ambiguous or ill‐defined, or when the input curves are disconnected, the surfacing stage cannot be robustly applied. Figures 15, 19, and 21 show our output on the same input as Pan et al. [PLS*15]. Limitations of our method are discussed in Section 8.7.
Figure 15
Comparison with prior methods. Each row shows one sketch example: (top) toothpaste, (middle) mouse, (bottom) sewing machine. Columns compare reconstructions from VIPSS [HCJ19], Yu et al. [YAB*22], and our method. For the toothpaste example, Yu et al.'s method requires a manually provided proxy; we also applied a small perturbation to improve topology in this case.
Figure 21
A gallery of reconstruction results created with SharpEdge. Each model is shown in four stages: input sketch, optimized dual normals, initial reconstructed mesh, and optimized mesh. See the supplemental materials for an interactive gallery.

8.5.1 Disconnected Curve Networks

Our pipeline can operate on disconnected or partially disconnected curve networks (Figures 13 and 14). The behaviour of the reconstructed surface depends on how topological constraints are preserved. In particular, the two‐normal optimization benefits from constraints propagated through high‐valence junctions. When such junctions are preserved, the normal field remains well‐constrained even if portions of the curves are disconnected elsewhere.
Figure 13
Behaviour of our method on disconnected curve networks. Top row: Side curves are disconnected directly at lower valence‐4 junctions, removing critical topological constraints. The normal field becomes under‐constrained, causing the reconstructed surface to pinch inward. Bottom row: side curves are disconnected at interior locations while high‐valence junctions remain intact. In this case, the optimized normal field and reconstructed surface are visually indistinguishable from those produced from fully connected input.
Figure 14
Even when sketches consists of multiple disconnected components, our method produces reasonable results, provided that high‐valence vertex constraints are preserved (Figure 20). Compare this output to the connected curve networks in Figures 2 and 7. VIPSS produces overly smooth results whether or not the curve networks are connected.
We illustrate this behaviour with several examples. In the spherecylinder model (Figure 13), disconnecting the four side curves of the cylinder near the middle, keeping the high‐valence junctions at the top and bottom intact, produces a two‐normal field and reconstructed surface that are visually indistinguishable from the fully connected case. In contrast, disconnecting the side curves directly from the lower valence‐4 junctions removes critical constraints, causing the two‐normal field to become under‐constrained. The resulting surface pinches inward, exhibiting behaviour similar to that of a minimal surface. Such outcomes reflect ambiguity in the input curve network rather than failure of the optimization. Figure 14 provides additional examples of disconnected curve networks and compares the resulting surfaces produced by our method and VIPSS.
This behaviour highlights an assumption made by our approach: topological connectivity at junctions meaningfully encodes the designer's intent. This assumption is shared by prior cycle‐based methods such as Zhuang et al. [ZZCJ13] and Pan et al. [PLS*15], which likewise rely on clean, connected curve networks. In contrast, methods such as VRSurf, Yu et al. [YAB*22], VIPSS, and NeuVAS do not require topological connectivity and can operate on messy input. However, some approaches produce smooth surfaces by design (or, in the case of NeuVAS, require manual annotation to introduce sharp features).
A large body of prior work addresses the clean‐up and consolidation of 2D sketches by merging duplicate strokes, repairing junctions, and removing small inconsistencies. Developing analogous preprocessing techniques for 3D curve networks remains an open problem. Such curve‐cleanup algorithms could be applied prior to our pipeline to improve robustness to disconnected or noisy input, and we consider this a promising direction for future work.

8.6 Ablation

With only one normal per edge, the surface reconstruction is fundamentally limited. Sharp features are smoothed out, particularly in sparse curve networks where neighbouring edges provide insufficient directional cues. The surface further exhibits bulging artifacts in regions with weak constraints, and thin structures could easily be lost since a single normal cannot capture both sides of a sheet (Figure 17).
Figure 17
Comparison of initial surfacing with one normal per edge (left) and two normals per edge (right) on the mouse example. Single‐normal reconstruction oversmoothes sharp features. The SHM isosurface bulges outward and is unable to interpolate the curves. Our automatically computed dual normals better preserve the intended structure.

8.7 Limitations and Future Work

The main limitation of our approach lies in the surface generation stage. The initial surfaces we reconstruct from the signed heat method have two drawbacks: (i) it may produce a surface that is not a valid manifold, and (ii) it can occasionally miss very short or densely packed input edges. In the latter case, our pipeline can still proceed, but the reconstructed surface does not fully follow the input in feature regions, leading to a slight loss of fidelity. In the former case, non‐manifold outputs can break subsequent stages such as remeshing, such as in the vacuum where remeshing fails due to non‐manifoldness (Figure 18). Nevertheless, optimization can still be applied, and we observe that it often reduces artifacts and partially recovers missing regions. We show additional imperfect examples in Figure 20.
Figure 18
Recovering from a poor quality initial surface. (a) This vacuum example exhibits two artifacts: a ‘bite’ where a curve edge is missed (on the right) and non‐manifold edges and vertices after zero‐length edges are collapsed (red highlights). These artifacts prevent remeshing. (b) The mesh can still be cut along the other curve edges, allowing our pipeline to continue. (c) Optimization without remeshing reduces the “bite” and improves the surface quality, though the mesh element quality remains limited. (d) A further optimization iteration (recalculating the discrete Laplace operator) makes the ‘bite’ disappear.
Figure 20
Examples exhibiting minor artifacts. Even when remeshing fails or short edges are missed, we can still generate a surface, although some regions exhibit a slight loss of fidelity.
We also demonstrate this limitation on the espresso machine (Figure 19). The espresso machine consists of many individual components and is highly concave. In this example, we apply our method to the largest component only. Overall, the dual normal field reconstructed by our method remains largely consistent at a global scale and captures the main shape of the object well. However, due to the geometric complexity of this model, some local artifacts remain, particularly in regions such as the upper‐front bulge and the connection between the base to the upper body (see arrows in Figure 19a). Locally inconsistent normal orientations lead to distorted surface regions in the initial reconstruction (Figure 19b), reflecting the limitation of the initial surface extraction stage. Despite these artifacts, the pipeline can still generate a surface and proceed with optimization. Repeated application of our surface optimization reduces artifacts and improves overall surface quality (Figure 19c, d), similar to the behaviour observed in the vacuum example.
Figure 19
The espresso machine is a geometrically complex example. We demonstrate our method on its largest component. (a) The reconstructed normal field exhibits local artifacts (see arrows). (b) The initial reconstructed surface contains noticeable artifacts, partly due to locally inconsistent normals and limitations of the initial surface extraction stage. (c) The surface after one optimization pass. (d) A further optimization pass improves the overall surface quality, exhibiting behaviour similar to that observed in Figure 18. These results suggest that a more accurate initial surface extraction could further improve reconstruction quality.
These limitations suggest that a more robust surface extraction approach is needed. Unfortunately, we are not able to use the better‐studied marching cubes, since we require an unstructured tetrahedral mesh to conform to our input curves. The remedies we explored (§6) are insufficiently robust. To the best of our knowledge, existing methods do not provide a suitable solution in this setting.
Our approach could be improved with a small amount of user input. Our dual‐normal adaptation of the signed heat method matches convex sharp features and rounds concave sharp features (Figure 8). Users could click to flip the interpretation.
Finally, our framework inherits limitations from the input curve networks themselves. Inaccurate or disconnected curves can reduce the effectiveness of our constraints. We can handle disconnected curves with the pairwise weight (omitting parallel transport). However, 3D sketch input with disconnected curves often also contains messy strokes, in which curves do not meet precisely at junctions. Naively connecting or snapping nearby sketches would produce strokes which overshoot junctions, leading to non‐manifold edges. We wish to explore approaches to neatening messy sketches akin to this active research area in 2D.

9 Conclusion

We presented SharpEdge, an algorithm for reconstructing smooth surfaces with sharp features from sparse (possibly disconnected) curve networks. Our two‐normal optimization approach automatically discovers bimodal normals along all edges. Our approach is based on a notion of incidence and continuity at high‐valence junctions. By adapting the signed heat method to curve elements, we are able to recover an initial surface suitable for mesh and surface refinement. Our cut‐remesh‐optimize approach provides precise boundary conditions for Stanko et al.'s [SHBSS16] optimization approach, which ties together information we have exactly (normals along patch boundaries) with vertex positions. The resulting surfaces are high quality while interpolating the input curves with intended sharpness. The most brittle stage of our approach is isosurface extraction.

Acknowledgements

We are grateful to Nicole Feng, Alec Jacobson, Misha Kazhdan, and Mario Botsch for valuable discussions. We also thank Jialin Huang for running the iPSR baseline. Authors Yu and Gingold were supported in part by the United States National Science Foundation (IIS‐2402893). Sheffer acknowledges the support of the Natural Sciences and Engineering Research Council of Canada (grant RGPIN‐2024‐03981).

Ethics Statement

This study did not involve human participants, animals, or sensitive data.

Affiliations

1George Mason University, Fairfax, Virginia, USA
2University of British Columbia, Vancouver, Canada
3Adobe Research, San Jose, California, USA

ORCID

X. Yu https://orcid.org/0000-0002-3748-8295
Y. Gingold https://orcid.org/0000-0002-5381-2104

Footnotes

1
If ∥n∥<ε$\Vert \mathbf {n}\Vert < \varepsilon$, we instead use the average incident unit edge vectors—the graph Laplacian—as an approximation of the discrete mean curvature normal Hn$\mathbf {H}_n$. If that vector also vanishes, e.g., in flat or saddle‐like regions, we instead apply PCA to the set of incident edge vectors to estimate a local tangent plane.

References

Abbasinejad F., Joshi P., Amenta N.: Surface Patches from Unorganized Space Curves. URL: https://doi.org/10.1145/2261250.2261314.
2013
Abbasinejad F., Joshi P., Grimm C., Amenta N., Simons L.: Surface patches for 3D sketching. In Proceedings of the International Symposium on Sketch‐Based Interfaces and Modeling (New York, NY, USA, July 2013), SBIM '13, Association for Computing Machinery, pp. 53–60. URL: https://dl.acm.org/doi/10.1145/2487381.2487387.
2018
Barill G., Dickson N., Schmidt R., Levin D. I., Jacobson A.: Fast winding numbers for soups and clouds. ACM Transactions on Graphics 37, 4 (2018), 43:1–43:12. URL: https://doi.org/10.1145/3197517.3201337.
1975
Bishop R. L.: There is More than One Way to Frame a Curve. The American Mathematical Monthly 82, 3 (Mar. 1975), 246–251. URL: https://doi.org/10.1080/00029890.1975.11993807.
2010
Botsch M., Kobbelt L., Pauly M., Alliez P., Lévy B.: Polygon Mesh Processing, 1 ed. A K Peters/CRC Press, 2010. URL: https://www.pmp‐book.org.
2008
Bergou M., Wardetzky M., Robinson S., Audoly B., Grinspun E.: Discrete elastic rods. In ACM SIGGRAPH 2008 Papers (New York, NY, USA, 2008), SIGGRAPH '08, Association for Computing Machinery. URL: https://doi.org/10.1145/1399504.1360662.
2024
Coiffier G., Béthune L.: 1‐Lipschitz Neural Distance Fields. Computer Graphics Forum 43, 5 (2024), e15128. URL: https://doi.org/10.1111/cgf.15128.
1973
Douglas D. H., Peucker T. K.: Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. The Canadian Cartographer 10, 2 (1973), 112–122. URL: https://doi.org/10.3138/FM57‐6770‐U75U‐7727.
2024
Erler P., Fuentes‐Perez L., Hermosilla P., Guerrero P., Pajarola R., Wimmer M.: PPSurf: Combining Patches and Point Convolutions for Detailed Surface Reconstruction. Computer Graphics Forum 43, 1 (2024), e15000. URL: https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.15000.
2020
Erler P., Guerrero P., Ohrhallinger S., Mitra N. J., Wimmer M.: Points2Surf Learning Implicit Surfaces from Point Clouds. In Computer Vision – ECCV 2020 (Cham, 2020), vol. 12350, Springer International Publishing, pp. 108–124. URL: https://link.springer.com/10.1007/978‐3‐030‐58558‐7_7.
2024
Feng N., Crane K.: A heat method for generalized signed distance. ACM Transactions on Graphics 43, 4 (2024), 92:1–92:19. URL: https://doi.org/10.1145/3658220.
2023
Feng Y.‐F., Shen L.‐Y., Yuan C.‐M., Li X.: Deep Shape Representation with Sharp Feature Preservation. Computer‐Aided Design 157 (Apr. 2023), 103468. URL: https://www.sciencedirect.com/science/article/pii/S0010448522002019.
2017
Gori G., Sheffer A., Vining N., Rosales E., Carr N., Ju T.: Flowrep: Descriptive curve networks for free‐form design shapes. ACM Trans. Graph. 36, 4 (2017). URL: https://doi.org/10.1145/3072959.3073639.
2019
Huang Z., Carr N., Ju T.: Variational implicit point set surfaces. ACM Trans. Graph. 38, 4 (July 2019). URL: https://doi.org/10.1145/3306346.3322994.
1995
Hanson A. J., Ma H.: Parallel Transport Approach to Curve Framing. Computer Science Technical Reports TR425, Indiana University, Feb. 1995. URL: https://scholarworks.iu.edu/dspace/items/39fbe931‐c2c8‐43a1‐9c9d‐a207b92d551e/full.
2025
Härenstam‐Nielsen L., Sang L., Saroha A., Araslanov N., Cremers D.: DiffCD: A Symmetric Differentiable Chamfer Distance for Neural Implicit Surface Fitting. In Computer Vision – ECCV 2024 (Cham, 2025), vol. 15131, pp. 432–447. URL: https://link.springer.com/10.1007/978‐3‐031‐73464‐9_26.
2022
Hou F., Wang C., Wang W., Qin H., Qian C., He Y.: Iterative Poisson Surface Reconstruction (iPSR) for Unoriented Points. ACM Transactions on Graphics 41, 4 (July 2022), 1–13. arXiv:2209.09510 [cs]. URL: http://arxiv.org/abs/2209.09510.
2018
Jacobson A., Panozzo D., et al.: libigl: A simple C++ geometry processing library, 2018. URL: https://libigl.github.io/.
2006
Kazhdan M., Bolitho M., Hoppe H.: Poisson surface reconstruction. In Eurographics Symposium on Geometry Processing (SGP) (Goslar, Germany, 2006), SGP '06, Eurographics Association, pp. 61–70. URL: https://doi.org/10.2312/SGP/SGP06/061‐070.
2025
Liu W., Li J., Chen X., Hou F., Xin S., Wang X., Wu Z., Qian C., He Y.: Diffusing Winding Gradients (DWG): A Parallel and Scalable Method for 3D Reconstruction from Unoriented Point Clouds. ACM Trans. Graph. 44, 2 (Apr. 2025), 19:1–19:18. URL: https://dl.acm.org/doi/10.1145/3727873.
2015
Pan H., Liu Y., Sheffer A., Vining N., Li C.‐J., Wang W.: Flow aligned surfacing of curve networks. ACM Transactions on Graphics 34, 4 (July 2015), 1–10. URL: https://dl.acm.org/doi/10.1145/2766990.
2019
Rosales E., Rodriguez J., Sheffer A.: Surfacebrush: From virtual reality drawings to manifold surfaces. ACM Trans. Graph. 38, 4 (2019). URL: https://doi.org/10.1145/3306346.3322970.
2020
Sharp N., Crane K.: A Laplacian for Nonmanifold Triangle Meshes. Computer Graphics Forum 39, 5 (Aug. 2020), 69–80. URL: https://onlinelibrary.wiley.com/doi/10.1111/cgf.14069.
2025
Scrivener D., Cui D., Coldren E., Abulnaga S. M., Bessmeltsev M., Chien E.: Faraday cage estimation of normals for point clouds and ribbon sketches. ACM Trans. Graph. 44, 4 (July 2025). URL: https://doi.org/10.1145/3731212.
2016
Stanko T., Hahmann S., Bonneau G.‐P., Saguin‐Sprynski N.: Smooth Interpolation of Curve Networks with Surface Normals. Eurographics Short Papers (2016), 4 pages. URL: https://diglib.eg.org/handle/10.2312/egsh20161005.
2015
Si H.: TetGen, a Delaunay‐Based Quality Tetrahedral Mesh Generator. ACM Trans. Math. Softw. 41, 2 (Feb. 2015), 11:1–11:36. URL: https://dl.acm.org/doi/10.1145/2629697.
2025
Sureshkumar A., Parakkat A. D., Bonneau G.‐P., Hahmann S., Cani M.‐P.: Vrsurf: Surface creation from sparse, unoriented 3d strokes. Computer Graphics Forum 44, 2 (2025), e70071. URL: https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.70071.
2007
Sullivan J. M.: Curvatures of Smooth and Discrete Surfaces, Oct. 2007. arXiv:0710.4497 [math]. URL: http://arxiv.org/abs/0710.4497.
2025
Wang P., Dong Q., Liang F., Pan H., Yang L., Zhang C., Lin G., Zhang C., Zhou Y., Tu C., Xin S., Sheffer A., Li X., Wang W.: NeuVAS: Neural implicit surfaces for variational shape modeling. ACM Transactions on Graphics (SIGGRAPH Asia) (2025). URL: https://doi.org/10.1145/3763331.
2023
Wang Z., Zhang Y., Xu R., Zhang F., Wang P.‐S., Chen S., Xin S., Wang W., Tu C.: Neural‐Singular‐Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian. ACM Transactions on Graphics 42, 6 (Dec. 2023), 1–14. URL: https://dl.acm.org/doi/10.1145/3618311.
2014
Xu B., Chang W., Sheffer A., Bousseau A., McCrae J., Singh K.: True2Form: 3D curve networks from 2D sketches via selective regularization. ACM Transactions on Graphics 33, 4 (July 2014), 1–13. URL: https://dl.acm.org/doi/10.1145/2601097.2601128.
2023
Xu R., Dou Z., Wang N., Xin S., Chen S., Jiang M., Guo X., Wang W., Tu C.: Globally consistent normal orientation for point clouds by regularizing the winding‐number field. ACM Trans. Graph. 42, 4 (jul 2023). URL: https://doi.org/10.1145/3592129.
2025
Xia J., Ju T.: Variational surface reconstruction using natural neighbors. ACM Trans. Graph. 44, 4 (July 2025). URL: https://doi.org/10.1145/3731191.
2022
Yu E., Arora R., Bærentzen J. A., Singh K., Bousseau A.: Piecewise‐smooth surface fitting onto unstructured 3d sketches. ACM Trans. Graph. 41, 4 (July 2022). URL: https://doi.org/10.1145/3528223.3530100.
2021
Yu E., Arora R., Stanko T., Bærentzen J. A., Singh K., Bousseau A.: CASSIE: Curve and surface sketching in immersive environments. In ACM Conference on Human Factors in Computing Systems (CHI) (2021). URL: https://doi.org/10.1145/3411764.3445158.
2021
Yu X., DiVerdi S., Sharma A., Gingold Y.: Scaffoldsketch: Accurate industrial design drawing in vr. In ACM Symposium on User Interface Software and Technology (2021), UIST '21, Association for Computing Machinery, pp. 372–384. URL: https://doi.org/10.1145/3472749.3474756.
2013
Zou M., Ju T., Carr N.: An algorithm for triangulating multiple 3d polygons. In Proceedings of the Eleventh Eurographics/ACM SIGGRAPH Symposium on Geometry Processing (SGP) (Goslar, Germany, 2013), SGP '13, Eurographics Association, pp. 157–166. URL: https://doi.org/10.1111/cgf.12182.
2025
Zhang Y., Wang S., Bessmeltsev M.: Variational neural surfacing of 3d sketches. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers (New York, NY, USA, 2025), SA Conference Papers '25, Association for Computing Machinery. URL: https://doi.org/10.1145/3757377.3763900.
2013
Zhuang Y., Zou M., Carr N., Ju T.: A general and efficient method for finding cycles in 3d curve networks. ACM Trans. Graph. 32, 6 (2013), 180:1–180:10. URL: https://doi.org/10.1145/2508363.2508423.