Research map / Point clouds / Shape and surface completion
Toward Robust Neural Reconstruction from Sparse Point Sets
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 unsupervised neural SDFs from sparse, noisy unoriented point clouds by applying distributionally robust optimization (Wasserstein and Sinkhorn dual forms) over query distributions on top of Neural Pull.
Problem
Learning accurate signed distance functions from sparse and noisy point clouds without ground-truth supervision, where existing unsupervised methods overfit and produce holes and hallucinations.
Approach
Replaces the empirical query distribution in Neural Pull with the worst-case distribution inside a Wasserstein ball, using the dual form of DRO; an entropic (Sinkhorn) regularized variant samples diffuse adversarial queries from a Gaussian around each query, which converges faster and improves robustness.
Key results
- On ShapeNet (Table, Chair, Lamp) from 1024 noisy points, SDRO reaches CD1 0.63 and CD2 0.012 versus NAP's 0.76 and 0.020.
- On 3D Scene, SDRO achieves mean CD1 0.020 and NC 0.881, improving on NP (mean CD1 0.097) and NAP (0.041).
- On Faust, NAP (CD1 0.220) slightly outperforms SDRO (0.251), so the method is not uniformly best, especially with clean dense inputs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · ShapeNet | CD1: 0.77 -> 0.63 | Helped |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · ShapeNet | CD2: 0.015 -> 0.012 | Helped |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · ShapeNet | NC: 0.87 -> 0.90 | Helped |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · Faust | CD1: 0.255 -> 0.251 | Helped |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · Faust | CD2: 0.002 -> 0.002 | No clear effect |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · 3D Scene mean | CD1: 0.022 -> 0.020 | Helped |
| replace the Wasserstein distance with the Sinkhorn distance (ablation) | SDF WDRO to form SDF SDRO · 3D Scene mean | NC: 0.872 -> 0.881 | Helped |
| use SDF SDRO instead of NP (ablation) | SDF reconstruction · ShapeNet Tables, σ = 0.0 | CD1: 0.73 -> 0.43 | Helped |
| use SDF SDRO instead of NP (ablation) | SDF reconstruction · ShapeNet Tables, σ = 0.005 | CD1: 1.07 -> 0.65 | Helped |
| use SDF SDRO instead of NP (ablation) | SDF reconstruction · ShapeNet Tables, σ = 0.025 | CD1: 2.45 -> 1.54 | Helped |
| use SDF SDRO instead of NAP | SDF reconstruction · Faust | CD1: 0.220 -> 0.251 | Hurt |
| use SDF SDRO instead of NAP | SDF reconstruction · Faust | NC: 0.956 -> 0.955 | No clear effect |
Surprising findings
- On clean, dense Faust inputs, NAP performs better than the proposed method, which the authors attribute to NAP's adaptive local radii.
Limitations
- Underperforms NAP on clean, dense point clouds (e.g., Faust).
- WDRO converges slowly and its worst-case distribution is discrete.
- Hyperparameters (lambda, rho, Ns) tuned on the SRB benchmark.
Experimental setup
Datasets: ShapeNet Faust 3D Scene SemanticPOSS BlendedMVS Tanks and Temples
Benchmarks: ShapeNet Faust 3D Scene SemanticPOSS BlendedMVS Tanks and Temples
Metrics: CD1 CD2 F-Score Normal Consistency
Baselines: NP NAP SparseOcc NTPS POCO CONet On-Surf NKSR SPSR OG-INR GridPull SAP
Builds on: Neural Pull NAP (Neural Adversarial Pull) Wasserstein distributionally robust optimization dual form Sinkhorn distance DRO
Closest papers by meaning
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (NeurIPS 2024 · similarity 0.89)
- Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries (ICML 2024 · similarity 0.87)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.87)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.86)
- Unsupervised Occupancy Learning from Sparse Point Cloud (CVPR 2024 · similarity 0.86)
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.85)
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.