Abstract
Voronoi diagrams are highly compact representations that are used in various Graphics applications. In this work, we show how to embed a differentiable version of it - via a novel deep architecture - into a generative deep network. By doing so, we achieve a highly compact latent embedding that is able to provide much more detailed reconstructions, both in 2D and 3D, for various shapes. In this tech report, we introduce our representation and present a set of preliminary results comparing it with recently proposed implicit occupancy networks.
Introduction
Representing 3D geometry for deep learning typically involves a trade-off between explicit methods (meshes, point clouds) and implicit methods (MLPs representing occupancy or distance fields). While implicit models like Occupancy Networks (OccNet) offer continuous representations, they rely on global support—where a single weight change can impact the entire shape. This lack of locality often necessitates large, over-parameterized networks to capture fine details.
VoronoiNet addresses this by introducing a differentiable Voronoi diagram layer. By partitioning space into local cells, the model achieves high-fidelity reconstruction with extreme parameter efficiency. Unlike hybrid methods using convex primitives, VoronoiNet guarantees non-overlapping regions, making it ideal for downstream tasks like physical simulation and collision detection.
Methodology: Differentiable Voronoi Diagrams
The core of VoronoiNet is the representation of a shape as a collection of Voronoi cells. Each cell is defined by a "site" and is pre-assigned a binary label: "inside" (1) or "outside" (0). The final shape is the union of all "inside" cells.
To make this representation trainable via backpropagation, the authors replace the non-differentiable argmin operation used in standard Voronoi diagrams with a soft-argmin function. This allows gradients to flow from the reconstruction loss back to the site coordinates. The pipeline consists of an encoder that maps input data to a latent code, and a decoder that translates this code into $K$ site positions.
To ensure stable training and high-quality tessellation, the model employs specific regularization terms:
- Centroidal Voronoi Loss: Encourages sites to move toward the center of their cells.
- Signed Distance Loss: Ensures "inside" sites remain within the ground-truth boundaries.
Experimental Evaluation
The model was evaluated on its ability to overfit a 3D sphere and its performance on the MNIST dataset.
Parameter Efficiency
A key advantage of VoronoiNet is its extreme compression capability. In tests on a synthetic 3D sphere, VoronoiNet achieved lower reconstruction error (Hausdorff distance) using only ~1,000 parameters compared to an OccNet model requiring ~800,000 parameters. This massive reduction highlights the efficiency of local support.
MNIST Reconstruction and Interpolation
On the MNIST dataset, VoronoiNet with 128 cells outperformed OccNet models with 512 neurons and matched the performance of models with 16,000 neurons. Furthermore, the latent space exhibits meaningful topological organization. Linear interpolation between latent codes results in smooth transformations, such as a "3" morphing into an "8" and then a "5," reflecting the geometric similarities between these digits.
Conclusion and Future Work
VoronoiNet demonstrates that embedding geometric priors like Voronoi diagrams into deep networks can lead to highly efficient and interpretable shape representations. While currently focused on 2D and simple 3D shapes, future work aims to scale the method to complex datasets like ShapeNet and explore its utility in high-dimensional data and physical simulations.
Summary updated