Learning Continuous Mesh Representation with Spherical Implicit Surface

Zhongpai Gao

Introduction

In computer graphics and 3D vision, the representation of geometric shapes is a fundamental challenge. Traditionally, 3D objects are stored as discrete meshes—collections of vertices and triangular faces. While effective for rendering, these discrete structures are inherently resolution-dependent. Increasing the level of detail often requires complex interpolation techniques that can smooth over sharp features or introduce artifacts.

This paper introduces Spherical Implicit Surface (SIS), a novel framework that reimagines 3D meshes as continuous functions rather than discrete point sets. By mapping 3D geometry onto a spherical coordinate system and utilizing deep neural networks to learn this mapping, the authors enable high-fidelity reconstruction and arbitrary-resolution upsampling for complex shapes like human faces and bodies.

The Limitations of Current Representations

Most existing 3D representations fall into two categories: discrete meshes or volumetric implicit functions.

  1. Discrete Meshes: These are the industry standard but suffer from fixed topology and resolution. Upsampling a low-poly mesh usually relies on simple geometric interpolation, which fails to recover high-frequency details like wrinkles or muscle definitions.
  2. Volumetric Implicit Functions: Methods such as Signed Distance Functions (SDFs) or Occupancy Fields represent shapes within a 3D volume. While continuous, they are computationally expensive to train and require post-processing algorithms like Marching Cubes to extract a usable surface mesh, which can be slow and memory-intensive.

SIS addresses these issues by providing a surface-direct continuous representation that is both computationally efficient and capable of capturing fine geometric details.

Methodology: The Spherical Implicit Surface

The core innovation of SIS is treating a 3D mesh as a continuous function: $v = f(\theta, \phi)$, where $(\theta, \phi)$ represents coordinates on a unit sphere.

1. Spherical Parameterization

Leveraging the uniformization theorem, the authors observe that many common 3D objects (or their components) are topologically equivalent to a sphere (genus-0). These shapes can be "unwrapped" onto a spherical coordinate system. For more complex topologies, such as a face with separate eye sockets, the model decomposes the template into multiple genus-0 parts to maintain the mapping's integrity.

2. The SIS Decoder and Fourier Features

The mapping function $f$ is implemented via a Multilayer Perceptron (MLP). However, standard MLPs often suffer from "spectral bias," where they struggle to learn high-frequency details, resulting in overly smooth surfaces. To counteract this, the authors apply Fourier feature mapping to the input spherical coordinates. This allows the network to capture sharp edges and intricate surface textures that would otherwise be lost.

3. Conditioning for Reconstruction and Super-Resolution

The model is designed to be flexible, supporting two primary modes of operation:

  • Global Conditioning: For shape reconstruction, a mesh encoder (LSA-Conv) extracts a global feature vector. The decoder uses this vector to reconstruct the entire shape from the spherical domain.
  • Local Conditioning: For super-resolution tasks, the model utilizes local features distributed across the surface. A Feature Fusion Module employs barycentric interpolation to provide a continuous feature field. This allows the decoder to predict high-resolution geometry even when starting from sparse or noisy input points.

4. Self-Supervised Learning

The training process is self-supervised, utilizing a combination of L1 reconstruction loss to ensure geometric accuracy and a Laplacian regularization term. The regularization is critical for maintaining surface smoothness and preventing structural artifacts during the continuous mapping process.

Experimental Results

The authors validated SIS using two prominent 3D benchmarks: the COMA dataset (human facial expressions) and the DFAUST dataset (dynamic human body movements).

Mesh Reconstruction

On the COMA dataset, SIS demonstrated remarkable precision, achieving an L2 reconstruction error of 0.179 mm. This significantly outperformed previous continuous state-of-the-art methods like TDeform (0.946 mm) and performed competitively with specialized discrete convolutional models (0.172 mm), proving that a continuous representation does not have to sacrifice accuracy.

Super-Resolution

In super-resolution tests, the model was tasked with upsampling a full mesh from only 500 sparse points. SIS achieved an error of 3.321 mm on the DFAUST dataset, a substantial improvement over traditional Barycentric Interpolation (BCI), which yielded an error of 5.323 mm. This highlights the model's ability to "fill in the gaps" of sparse data using its learned continuous prior.

Conclusion and Future Directions

Spherical Implicit Surface (SIS) represents a significant step toward efficient, continuous 3D modeling. By transforming 3D geometry into a 2D spherical mapping problem, it bypasses the heavy computational costs of volumetric methods while maintaining the flexibility of continuous functions.

Despite its success, some challenges remain. The current requirement for manual decomposition of complex topologies into genus-0 parts suggests a need for automated topological partitioning. Additionally, while Fourier features are effective, exploring more advanced coordinate encodings (such as hash-based grids) could further enhance the model's ability to represent ultra-fine details and improve training speeds. As it stands, SIS provides a robust framework for high-resolution 3D shape generation and reconstruction.

Summary updated

BibTeX
@inproceedings{gao2023learning,
  author = {Zhongpai Gao},
  title = {{Learning Continuous Mesh Representation with Spherical Implicit Surface}},
  booktitle = {FG},
  year = {2023},
  url = {https://openreview.net/forum?id=zUz6nFHOfb},
}
Figure