Research map / Point clouds / Shape and surface completion
Unsupervised Occupancy Learning from Sparse Point Cloud
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.
Learns a neural occupancy field per shape from sparse, noisy, unoriented point clouds without ground-truth supervision, using margin-based uncertainty sampling and entropy polarization regularization.
Problem
Reconstructing watertight 3D shapes from sparse, noisy, unoriented point clouds in an unsupervised setting, where SDF-based methods struggle without ground-truth labels.
Approach
Learn a binary occupancy field instead of an SDF by supervising its decision boundary via a Newton-Raphson root-finding (margin uncertainty sampling) loss that pulls boundary points onto the input point cloud, plus an entropy polarization regularizer that minimizes entropy in space and maximizes it at input points.
Key results
- On ShapeNet sparse noisy inputs, CD1 0.76 vs 1.07 for the next best (G-Pull); CD2 0.020 vs 0.032.
- On Faust, CD1 0.260 vs 0.308 for POCO (best baseline), with FS 0.974 vs 0.981 for POCO.
- On 3D Scene mean CD1 0.027 and CD2 0.003, versus NTPS at 0.077 and 0.010.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn occupancy fields with margin uncertainty sampling and entropy regularization instead of the compared reconstruction baselines | Ours · ShapeNet | CD1: 1.07 -> 0.76 | Helped |
| learn occupancy fields with margin uncertainty sampling and entropy regularization instead of the compared reconstruction baselines | Ours · Faust | CD1: 0.308 -> 0.260 | Helped |
| learn occupancy fields with margin uncertainty sampling and entropy regularization instead of the compared reconstruction baselines | Ours · Faust | FS: 0.981 -> 0.974 | Hurt |
| learn occupancy fields with margin uncertainty sampling and entropy regularization instead of NTPS | Ours · 3D Scene mean | CD1: 0.077 -> 0.027 | Helped |
| add entropy regularization to margin uncertainty sampling (ablation) | Ours · ShapeNet Tables | CD1: 1.02 -> 0.77 | Helped |
| replace entropy regularization with margin-based regularization (ablation) | Ours · ShapeNet Tables | CD1: 0.77 -> 0.80 | Hurt |
| replace entropy regularization with SA-Occ regularization (ablation) | Ours · ShapeNet Tables | CD1: 0.77 -> 0.95 | Hurt |
| increase input density from sparse to dense (ablation) | Ours · SRB | CD1: 0.49 -> 0.20 | Helped |
| increase input noise std. dev. from 0 to 0.025 (ablation) | Ours · ShapeNet Tables | CD1: 0.56 -> 2.16 | Hurt |
Surprising findings
- The sign-agnostic occupancy supervision (SA-Occ) fails to converge without priors, though it helps as a regularizer.
Limitations
- Excess regularization can make reconstructions overly smooth, hurting normal consistency on some benchmarks.
- Requires per-shape MLP fitting with hyperparameter tuning per benchmark or object for best results.
Experimental setup
Datasets: ShapeNet Faust 3D Scene Surface Reconstruction Benchmark (SRB)
Benchmarks: ShapeNet Faust 3D Scene SRB
Metrics: CD1 CD2 Normal Consistency F-Score Hausdorff distance
Baselines: SPSR OG-INR N-Pull G-Pull NTPS DiGs NDrop SAP NSpline POCO CONet On-Surf
Builds on: N-Pull Gropp et al. Eikonal regularization (IGR) Atzmon et al. SAL POCO
Closest papers by meaning
- Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries (ICML 2024 · similarity 0.89)
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (NeurIPS 2024 · similarity 0.88)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.88)
- Learning Occupancy for Monocular 3D Object Detection (CVPR 2024 · similarity 0.87)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.86)
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.