HyperCube: Implicit Field Representations of Voxelized 3D Models (Student Abstract)

Magdalena Proszewska, Marcin Mazur, Tomasz Trzcinski, Przemyslaw Spurek

Abstract

Implicit field representations are a cornerstone of modern 3D shape generation, yet existing methods like IM-NET often suffer from high computational costs and structural artifacts such as "holes" in reconstructed meshes. We introduce HyperCube, a compact hypernetwork-based architecture designed to address these limitations. By extending the model to use interval arithmetic (HyperCube-Interval), we enable the classification of entire 3D volumes (voxels) simultaneously. This approach ensures structural integrity and significantly reduces training time and memory overhead. Our evaluations on the ShapeNet dataset demonstrate that HyperCube achieves competitive reconstruction quality while being substantially more efficient than previous state-of-the-art methods.

Introduction

Generating high-quality 3D shapes typically involves learning implicit fields—functions that determine whether a point in space lies inside or outside an object's boundary. While models like IM-NET have proven effective at creating smooth surfaces, they face three primary challenges: high training complexity, a requirement for global dataset knowledge that hinders scaling, and sampling artifacts. The latter often manifests as disconnected "empty spaces" or holes near object boundaries because the model only processes individual points sampled within voxels.

HyperCube addresses these issues by leveraging a hypernetwork paradigm to generate weights for a compact target network on the fly, decoupling model size from dataset complexity.

Methodology

The HyperCube architecture is proposed in two primary configurations:

  1. Standard HyperCube: An input feature vector representing a voxel is passed to a hypernetwork ($H_\phi$), which outputs the weights ($\theta$) for a small target network ($T_\theta$). This target network then performs the occupancy classification.
  2. HyperCube-Interval: This advanced version replaces the standard Multi-Layer Perceptron (MLP) in the target network with an IntervalNet. By employing interval arithmetic, the model processes a continuous 3D interval (the entire voxel) rather than discrete points.
Comparison of point-based sampling in standard HyperCube versus volume-based processing in HyperCube-Interval.

To optimize the interval-based model, we utilize a "worst-case accuracy" loss function. This ensures that the entire volume of a voxel is correctly classified, effectively regularizing the classification boundary and preventing the formation of disconnected fragments.

Experimental Results

We evaluated HyperCube using the ShapeNet dataset across five categories: Plane, Car, Chair, Rifle, and Table.

Efficiency and Scalability

A key advantage of the HyperCube architecture is its efficiency. As shown in our performance benchmarks, HyperCube maintains a significantly lower memory footprint and faster training time compared to IM-NET across various batch sizes.

Training time (s) and GPU memory usage (MiB) comparison between IM-NET and HyperCube as batch size increases.

Reconstruction Quality

HyperCube-Interval consistently produces cleaner meshes with fewer connected components—meaning the resulting models are solid objects rather than fragmented clouds. Quantitatively, HyperCube achieved a Mean Squared Error (MSE) of 4.37 in the "Car" category, outperforming IM-NET's 4.99. In generative tasks, the interval-based approach showed superior Intersection over Union (IoU) metrics, particularly for complex geometries like airplanes.

Conclusion

HyperCube represents a significant step forward in efficient 3D shape representation. By combining hypernetworks with interval arithmetic, we provide a solution that is not only faster and lighter than existing methods but also more robust against common reconstruction artifacts. Future work will explore scaling IntervalNet to higher-resolution voxel grids and further refining the integration of these "plausible" 3D models into real-time applications like gaming and virtual reality.

Summary updated

Abstract

Implicit field representations offer an effective way of generating 3D object shapes. They leverage an implicit decoder (IM-NET) trained to take a 3D point coordinate concatenated with a shape encoding and to output a value indicating whether the point is outside the shape. This approach enables the efficient rendering of visually plausible objects but also has some significant limitations, resulting in a cumbersome training procedure and empty spaces within the rendered mesh. In this paper, we introduce a new HyperCube architecture based on interval arithmetic that enables direct processing of 3D voxels, trained using a hypernetwork paradigm to enforce model convergence. The code is available at https://github.com/mproszewska/hypercube.
BibTeX
@inproceedings{proszewska2024hypercube,
  author = {Magdalena Proszewska and Marcin Mazur and Tomasz Trzcinski and Przemyslaw Spurek},
  title = {{HyperCube: Implicit Field Representations of Voxelized 3D Models (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30499},
}
Figure