Graph Neural Networks for Seam Detection

• The paper introduces a GNN-based approach that reformulates seam detection as a binary edge classification problem on a dual graph constructed from 3D mesh data.
• Graph neural networks such as GCN, GAT, GraphSAGE, and GIN are utilized to replicate artist seam styles while addressing severe class imbalance through weighted losses and post-processing techniques like DST and skeletonization.
• Experimental results demonstrate that GraphSeam achieves a close replication of semantic boundaries with reduced UV distortion, outperforming traditional distortion-minimization methods while offering practical scalability on high-resolution meshes.

Graph neural networks (GNNs) for seam detection represent a supervised, data-driven paradigm for identifying semantically meaningful cut locations on 3D meshes, particularly in the context of UV mapping. By formulating seam detection as an edge classification problem on suitably constructed mesh-derived graphs, GNN-based approaches are capable of replicating artist-provided seam styles and discovering boundaries that go beyond the capabilities of distortion-minimization methods alone. Recent frameworks such as GraphSeam leverage the representational power of GNNs and cascaded graph post-processing to produce UV layouts that balance semantic control with geometric distortion minimization (Teimury et al., 2020).

1. Mesh-to-Graph Representation

Seam detection on triangle meshes $M=(V,E,F)$ is recast as a graph learning problem through dual-graph construction. Each mesh edge $e\in E$ is encoded as a vertex $\tilde{e}$ in the dual graph $\hat{G} = (\hat{E}, \hat{E}')$. Dual edges are established between $\tilde{e}=(i,j)$ and $\tilde{e}'=(i,k)$ or $\tilde{e}'=(j,\ell)$ if the corresponding mesh edges share a vertex. Node features for each vertex $x_v \in \mathbb{R}^6$ comprise 3D position, normal, and optional Gaussian curvature. Edge features $x_{\tilde{(i,j)}}\in\mathbb{R}^{12}$ are constructed by concatenating the features of edge endpoints: $x_{\tilde{(i,j)}} = [x_i \Vert x_j]$. Augmented dual variants that remove ordering ambiguity yield negligible accuracy benefit. This dual-graph formalism provides a structure in which seam detection is reframed as binary edge (dual-vertex) classification.

2. GNN Architectures and Learning Pipeline

The core task is binary edge classification on the dual graph, i.e., predicting $e\in E$0 as the seam likelihood for each edge. Multiple GNN backbones have been evaluated for this purpose, each incorporating residual skip-connections per layer:

• Graph Convolutional Network (GCN): Updates operate via symmetrically normalized adjacency, $e\in E$1 with ReLU activations.
• Graph Attention Network (GAT): Employs attention-weighted message passing to adaptively aggregate information from neighbor nodes.
• GraphSAGE: Utilizes an LSTM-based aggregator for neighborhood message synthesis and concatenation at each layer.
• Graph Isomorphism Network (GIN): Employs multi-layer perceptrons to enable maximally powerful injective aggregation.

All architectures culminate in outputting per-edge seam probabilities. A weighted cross-entropy loss function is optimized to counter the pronounced class imbalance (<10% edges are seams), using $e\in E$2 for seams and $e\in E$3 otherwise. Training employs Adam optimizer with standard learning rates, early stopping (50 epochs patience), and 3 GNN layers of 64 hidden units. GAT utilizes multiple attention heads and dropout for regularization. No explicit seam-length or style-consistency losses are applied; style consistency is enforced implicitly through supervision using artist-provided ground-truth.

3. Data, Supervision, and Semantic Style Replication

GraphSeam is trained on the Autodesk Character Generator (CG) dataset, comprising approximately 100 procedurally generated humanoid meshes with expert-labeled seam annotations. Meshes are preprocessed via decimation to 10,000 faces (CG10000), balancing geometric detail and computational tractability. Two toolchains support artist-driven data creation: a decimation tool for variable mesh resolutions and an augmentation tool injecting vertex-level Gaussian noise to diversify training samples while preserving seam style. The binary ground-truth vector $e\in E$4 specifies seam presence for every edge, directly encoding artist-preferred seam semantics. Datasets are partitioned into training (93 meshes), validation (3), and test (3) splits. Training access is strictly limited to the relevant partitions.

4. Graph-Based Post-Processing and Refinement

Raw GNN predictions frequently exhibit artifacts such as fragmented seams, thickened multi-edge bands, or omit edges critical for UV distortion minimization. To address these, two cascaded post-processing steps are introduced:

4.1 Distortion-Minimizing Steiner Tree (DST)

Edges with $e\in E$5 (typically $e\in E$6) are selected to form initial seams and component shells. Face-level normalized distortion $e\in E$7 is computed, and edge saliency $e\in E$8 weights are assigned. Terminal sets in the dual graph are updated via an approximate Steiner tree algorithm: shortest-path distances between seam dual-vertices are computed, a minimum spanning tree is constructed on the resulting complete graph, and edges are mapped back to the mesh. This procedure preferentially reconnects seams along minimal-distortion paths, addressing missing cuts and reducing UV distortion.

4.2 Skeletonization

To enforce 1D seam regularity, an iterative thinning ("skeletonization") algorithm reduces >1-edge-thick seam regions. Candidate seam vertices $e\in E$9 (where $\tilde{e}$0, with $\tilde{e}$1–$\tilde{e}$2) are pruned: vertices with lowest seam confidence are removed if shell connectivity and proximity constraints are preserved (using distance parameter $\tilde{e}$3–4). Small shells (size $\tilde{e}$4 for $\tilde{e}$5) are eliminated. The result is a refined, topological 1D skeleton corresponding to plausible UV seams.

5. Quantitative and Qualitative Evaluation

GraphSeam and its variants (GAT, GraphSAGE, GCN, GIN), augmented with DST and skeletonization, have been benchmarked against Autocuts [Poranne et al. ’17] and OptCuts [Li et al. ’18]. Evaluation on three held-out meshes (Ken, Sibilla, Xena) employs metrics such as false-positive rate (FPR), true-positive rate (TPR), overall accuracy (Acc), average per-face distortion (BPP, APP), and shell count:

Method	FPR (%)	TPR (%)	Acc (%)	Dist. BPP	Dist. APP	# Shells BPP	# Shells APP
Prop-GAT	0.34	99.00	99.63	0.524	0.135	30.6	68
Prop-GS	0.04	98.56	99.90	0.424	0.125	9	36
Prop-GCN	3.62	87.36	96.01	1.493	0.310	280	385
Prop-GIN	0.85	71.74	98.05	3.829	0.371	36	89

• Autocuts/OptCuts (without user input) yield single-shell UV maps with low distortion but high semantic error.
• GraphSeam+DST recovers semantic shells in a fully automatic fashion, reducing distortion close to OptCuts but at the cost of additional shells.
• Skeletonization reduces multi-edge seams to thin, plausible boundaries (e.g., GCN-DST: 385 → 19 shells).
• Qualitative results reveal close match to artist-labeled semantic boundaries, with meaningful anatomical separation.

6. Strengths, Limitations, and Open Research Directions

GraphSeam’s supervised GNN approach achieves automatic seam-style replication, scales to high-resolution meshes (10,000–100,000 faces), and is capable of semantic boundary discovery beyond distortion-based baselines. Combined post-processing fosters UV layouts with reduced distortion and regular seam geometry.

Limitations include the propensity for excessive fragmentation when post-processing thresholds are not optimal, absence of an end-to-end distortion loss term (necessitating reliance on post-hoc distortion correction), and sensitivity to hyperparameter choices such as $\tilde{e}$6, $\tilde{e}$7, and $\tilde{e}$8. Notably, style consistency is not explicitly regularized with loss terms; it emerges through ground-truth replication.

Emerging research directions include: incorporating differentiable UV-distortion proxies in the GNN loss, style-consistency regularization, multi-task architectures for joint seam and layout prediction, adaptive thresholding, and interactive artist-in-the-loop refinement.

7. Context and Comparative Overview

Prior to supervised graph learning approaches for seam detection, energy-minimization methods such as Autocuts and OptCuts formalized seam estimation via optimization of geometric criteria. However, they lacked the capability to encode semantic/artistic preferences. The adoption of GNNs, leveraging graph-structured representations and message passing, inaugurated a fundamentally new approach—enabling seam selections that generalize artist-provided stylistic tendencies and, via post-processing, approach or surpass conventional distortion minimization techniques in UV layout fidelity (Teimury et al., 2020). This framework supports the definition of seam styles at scale, and presents a platform for future integration of richer geometric and semantic priors.