Research map / 3D vision / Face and shape
HyperCube: Implicit Field Representations of Voxelized 3D Models (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HyperCube, a hypernetwork that generates weights of a small implicit decoder to classify whole voxels as inside/outside, plus a HyperCube-Interval variant using IntervalNet, for faster and more compact 3D implicit shape reconstruction.
Problem
Implicit field decoders such as IM-NET are cumbersome to train across many object classes and, because they only sample points inside voxels, produce classification-boundary artifacts and empty spaces in rendered meshes.
Approach
A hypernetwork conditioned on a voxel feature vector outputs the weights of a compact target network that predicts inside/outside labels; replacing the MLP target with interval arithmetic (IntervalNet) lets the model process entire voxel cubes instead of sampled points, removing empty-space artifacts.
Key results
- HyperCube reaches reconstruction results comparable to IM-NET on Plane, Car, Chair, Rifle, Table (e.g., Car CD 3.36 vs 5.44 for IM-NET, x10^-4), while training faster with lower GPU memory.
- In generative evaluation, HyperCube improves mean IoU on Car (86.34 vs 77.36) and Rifle (59.80 vs 51.53) over IM-NET.
- HyperCube-Interval produces fewer connected components (fewer disconnected pieces and empty-space artifacts) for Airplane and Car classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace IM-NET with HyperCube | 3D shape reconstruction · Plane, Car, Chair, Rifle, and Table | MSE: Plane: 2.14 -> 2.44; Car: 4.99 -> 4.37; Chair: 11.43 -> 9.07; Rifle: 1.91 -> 1.91; Table: 10.67 -> 9.37 | Mixed |
| replace IM-NET with HyperCube | 3D shape reconstruction · Plane, Car, Chair, Rifle, and Table | IoU: Plane: 78.77 -> 65.35; Car: 89.26 -> 90.05; Chair: 65.65 -> 72.61; Rifle: 72.88 -> 63.97; Table: 71.44 -> 73.78 | Mixed |
| replace IM-NET with HyperCube | 3D shape reconstruction · Plane, Car, Chair, Rifle, and Table | CD: Plane: 4.23 -> 4.74; Car: 5.44 -> 3.36; Chair: 9.05 -> 8.35; Rifle: 3.77 -> 4.20; Table: 11.54 -> 8.82 | Mixed |
| use interval arithmetic and IntervalNet instead of sampled points and an MLP target network (ablation) | HyperCube · Airplane and Car mesh reconstruction | number of connected components | Helped |
| use HyperCube rather than IM-NET | 3D shape generation · ShapeNet test-set objects | MSE, IoU, and CD | No clear effect |
Surprising findings
- HyperCube has lower IoU than IM-NET on the Plane class in the reconstruction table (65.35 vs 78.77).
Limitations
- Results are mixed: HyperCube underperforms IM-NET on some classes (e.g., Plane IoU, Chair MSE in generative setting).
- Evaluation limited to a subset of ShapeNet categories (100 test elements per category).
Experimental setup
Datasets: ShapeNet
Benchmarks: ShapeNet
Metrics: MSE IoU Chamfer Distance training time GPU memory number of connected components
Baselines: IM-NET
Builds on: IM-NET HyperCloud IntervalNet
Closest papers by meaning
- XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies (CVPR 2024 · similarity 0.87)
- GaussianCube: A Structured and Explicit Radiance Representation for 3D Generative Modeling (NeurIPS 2024 · similarity 0.85)
- GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation (ICLR 2025 · similarity 0.84)
- Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering (CVPR 2025 · similarity 0.83)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.83)
- MeshXL: Neural Coordinate Field for Generative 3D Foundation Models (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.