3DGeoMeshNet: 3D Mesh Reconstruction

• 3DGeoMeshNet is a graph-convolutional autoencoder that directly reconstructs 3D polygonal meshes by processing irregular graph data with anisotropic spatial convolutions.
• It employs a dual-path multi-scale encoder–decoder architecture that separately captures global shape properties and local geometric details through adaptive attention fusion.
• Empirical evaluations on the COMA facial mesh dataset show improved reconstruction accuracy under non-linear expression changes compared to earlier mesh processing methods.

Searching arXiv for the named paper and closely related mesh-network papers for citation support. 3D Geometric Mesh Network (3DGeoMeshNet) is a graph-convolutional autoencoder for reconstructing 3D polygonal meshes directly in mesh form. It is presented as a novel GCN-based framework that uses anisotropic convolution layers to effectively learn both global and local features directly in the spatial domain, rather than converting meshes into intermediate representations such as voxel grids or point clouds. Its core design is a multi-scale encoder–decoder with separate global and local pathways, together with attention-based fusion, and its reported evaluation on the COMA dataset containing human faces emphasizes reconstruction accuracy under strong non-linear expression changes (Nazir et al., 7 Jul 2025).

1. Problem setting and representational assumptions

3DGeoMeshNet targets 3D mesh reconstruction of human facial meshes, especially those with strong non-linear expression changes. The problem is framed around the non-Euclidean character of mesh data: vertices are irregularly connected, each vertex may have a different neighborhood pattern, and there is no consistent grid like in images. In that setting, standard CNNs are not directly applicable, and workaround representations such as voxelization and point clouds either lose geometric fidelity, require large memory and computation, or fail to preserve the original polygonal surface structure (Nazir et al., 7 Jul 2025).

The model treats the input mesh as $$(V_{\text{in}}, E_{\text{in}}, F_{\text{in}}, \mathbf{X}_{\text{in}})$$, where $V$ are vertices, $E$ edges, $F$ faces, and $\mathbf{X}\in \mathbb{R}^{N\times F}$ are vertex features. In practice, the paper uses normalized vertex positions plus mean principal curvature as an extra feature, so $F=4$. The curvature feature is defined by

$H(v)=\frac{1}{2}(k_1(v)+k_2(v)),$

where $k_1(v)$ and $k_2(v)$ are the principal curvatures at vertex $v$ (Nazir et al., 7 Jul 2025).

A further assumption is that meshes are aligned to a common template with fixed vertex count and topology during training, making graph-based reconstruction feasible and consistent. This constraint is central to the formulation: 3DGeoMeshNet reconstructs meshes in their original polygonal format, but it does so in a regime where connectivity is controlled rather than arbitrary (Nazir et al., 7 Jul 2025).

2. Dual-path multi-scale encoder–decoder

The architecture is an autoencoder in which an encoder maps input mesh features to a latent code $V$0, and a decoder reconstructs vertex features from that code. Its distinctive feature is a multi-scale dual-path design built around two complementary pathways. The global pathway captures large-scale shape and topology, uses graph convolution layers interleaved with down-sampling, and aggregates information across the whole mesh. The local pathway captures fine geometric details and local deformations, uses graph convolution layers with residual connections, and preserves more detailed vertex-level structure (Nazir et al., 7 Jul 2025).

In the encoder, a global encoder $V$1 and a local encoder $V$2 each produce a feature vector of size $V$3. These are concatenated and mapped to the latent vector $V$4. The global branch uses graph convolution layers plus down-sampling, followed by averaging vertex features across vertices to obtain a global descriptor. The local branch uses graph convolution layers with residual links, then flattens the resulting vertex features and linearly transforms them into a local descriptor (Nazir et al., 7 Jul 2025).

The decoder mirrors this decomposition. The latent vector is first split into $V$5 and $V$6, and the two branches reconstruct complementary feature maps. The global decoder uses global up-sampling plus graph convolutions, whereas the local decoder uses a linear layer plus several graph convolutions with residual connections. The fusion rule is

$V$7

so reconstruction is explicitly decomposed into global and local contributions (Nazir et al., 7 Jul 2025).

Multi-scale down-sampling and up-sampling are fixed to powers of two: each down-sampling reduces vertex count by a factor of 2, and each up-sampling doubles the number of vertices. This arrangement allows the global branch to see coarse shape structure while the local branch retains and reconstructs finer detail. A plausible implication is that the architecture is designed to separate topology-sensitive coarse information from expression-sensitive local variation without collapsing both into a single aggregation stream (Nazir et al., 7 Jul 2025).

3. Anisotropic spatial graph convolutions and adaptive fusion

A central technical choice is the use of anisotropic graph convolutions in the spatial domain. The graph convolution layers are implemented with Feature-Steered Graph Convolution (FeaStConv), followed by LeakyReLU. In this context, anisotropic convolution assigns different weights to different neighboring vertices depending on learned features or geometric context, unlike isotropic filters, which treat neighbors more uniformly (Nazir et al., 7 Jul 2025).

The paper positions this against two common limitations in earlier mesh GCNs. Spectral GCNs depend on Laplacian eigen-decomposition and typically use isotropic filters, which are described as not expressive enough for directional, anisotropic surface details. Spatial GCNs often rely on fixed neighbor orderings or local-only aggregation, which can be inconsistent across meshes and may not capture global topology well. FeaStConv is used precisely because it aggregates neighborhood features using learned dynamic weights on graph edges without requiring fixed neighbor ordering, spiral traversals, or spectral eigen-decomposition (Nazir et al., 7 Jul 2025).

Fusion between the global and local branches is not fixed. Unlike FaceCom, where the global/local blending weights were fixed scalars, 3DGeoMeshNet learns vertex-wise adaptive weights: $V$8 The attention module concatenates global and local decoded features at each vertex, uses two linear layers, applies ReLU in the first layer and Softmax in the second, and outputs per-vertex weights controlling how much each path contributes. This makes fusion adaptive: some vertices rely more on global context, others on local detail (Nazir et al., 7 Jul 2025).

This design clarifies a frequent misconception in mesh learning: direct mesh processing need not imply either spectral machinery or handcrafted neighbor ordering. In 3DGeoMeshNet, anisotropic spatial graph convolutions and attention-guided dual-path decoding are the mechanisms used to preserve polygonal mesh structure while handling both large-scale and local geometric variation (Nazir et al., 7 Jul 2025).

4. Latent space, loss functions, and reconstruction objective

The paper defines both reconstruction and latent-space regularization. The main reconstruction loss is mean squared error,

$V$9

where $E$0 is the Frobenius norm. This is combined with spherical regularization,

$E$1

and the total loss is

$E$2

The stated intent of the regularizer is to keep latent codes near a spherical manifold for smoother latent structure and better generalization (Nazir et al., 7 Jul 2025).

The latent representation is also used for interpolation and extrapolation. The paper gives

$E$3

for latent interpolation, and

$E$4

for latent extrapolation. These formulas situate 3DGeoMeshNet within the standard mesh-autoencoding tradition in which the latent code is not merely a compression vector but also a space for controlled geometric variation (Nazir et al., 7 Jul 2025).

The final reconstructed vertex positions are recovered by reversing normalization: $E$5 where $E$6 is the template mean and $E$7 is the scale or variance term used in normalization. This emphasizes that reconstruction is learned in normalized coordinates and then mapped back to the original geometric frame (Nazir et al., 7 Jul 2025).

5. Training protocol, dataset, and empirical results

The reported evaluation uses the COMA dataset: human facial meshes, 12 extreme expression classes, 12 subjects, 20,466 registered 3D meshes, and all aligned to a common template with $E$8 vertices. The split is 90% training and 10% testing, with 100 training samples used for validation. Training uses Adam with initial learning rate $E$9, the learning rate is halved every 50 epochs, batch size is 32, the model is trained for 300 epochs, the implementation is in PyTorch, the GPU is an NVIDIA GeForce RTX 4090, and training time is about 24 hours. Local convolution feature sizes are 8, 16, 32, and global convolution feature sizes are 32, 64, 128. Evaluation metrics are mean error in mm, median error in mm, and $F$0 error in mm (Nazir et al., 7 Jul 2025).

Selected reconstruction results reported in the paper are as follows:

Method	Mean error	Median error
3DGeoMeshNet ($F$1)	$F$2	$F$3
FaceCom (modified)	$F$4	$F$5
LSA-Conv	$F$6	$F$7

These numbers support the paper’s claim that 3DGeoMeshNet substantially outperforms most earlier methods in the comparison, while also noting that LSA-Conv still achieves slightly better error in the reported table (Nazir et al., 7 Jul 2025).

The ablation study is structurally important. Increasing latent size improves reconstruction, with reported progression from latent 32 at mean 0.831, latent 64 at 0.632, latent 128 at 0.522, and latent 256 at 0.516. Adding attention improves performance to 0.223; adding attention and residuals improves it to 0.177; adding attention, residuals, and curvature gives 0.171. Removing either pathway hurts strongly: local only gives mean 0.688 and global only gives mean 1.520. This validates the need for dual-scale fusion in the specific experimental setting (Nazir et al., 7 Jul 2025).

6. Position within mesh-network research and stated limitations

Within the broader literature, 3DGeoMeshNet belongs to the family of methods that operate directly on meshes rather than on converted representations. It differs, however, from several adjacent lines of work. MeshNet is a face-based neural architecture for 3D shape representation that uses faces as the basic processing unit, splits face features into spatial and structural components, and targets classification and retrieval rather than reconstruction (Feng et al., 2018). EMNN, by contrast, is a minimal extension of EGNN to mesh data that incorporates face messages and hierarchy while preserving $F$8-equivariance; it is explicitly framed around Euclidean symmetry and face-normal geometry rather than around autoencoding with global/local fusion (Trang et al., 2024).

The name should also be distinguished from 3DMeshNet, which addresses three-dimensional structured mesh generation for numerical simulation by embedding meshing-related differential equations into the loss function of neural networks and learning mappings from parametric to computational domains. That problem is unsupervised structured mesh generation, not polygonal mesh reconstruction of registered facial data (Peng et al., 2024). In another direction, a defense-related study proposes a 3DGeoMeshNet-style hybrid architecture in which a hierarchical, instance-based scene manager combines a traditional mesh scaffold with a neural visual representation such as 3D Gaussian Splatting; here the emphasis is geometric integrity for line-of-sight, route planning, and physics, rather than autoencoding of facial meshes (Kahl et al., 7 Nov 2025).

The limitations stated for 3DGeoMeshNet are specific. It is tested on non-textured, monochromatic meshes only; it assumes fixed vertex count and topology; it is evaluated primarily on human facial meshes; and it still relies on a common template registration (Nazir et al., 7 Jul 2025). These restrictions do not negate the reported results, but they delimit the scope of the architecture as presented. A plausible implication is that the model’s strongest claims concern registered mesh reconstruction under controlled topology, not arbitrary-topology mesh learning in the general case.