Research map / Point clouds / Shape and surface completion
MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step
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 MultiPull, an unsupervised overfitting method that learns multi-scale SDFs from a single raw point cloud by pulling query points coarse-to-fine using multi-level Fourier/Hadamard frequency features, with distance-aware, gradient-consistency and surface losses.
Problem
Reconstructing continuous surfaces from raw point clouds without ground-truth signed distances or normals; overfitting-based neural SDF methods over-smooth local details.
Approach
Map query points into a set of frequency features (Hadamard-product multiplicative filter network) and perform multi-step pulling where each step is conditioned on a different frequency level, with distance-aware attention weights, a min-over-steps gradient consistency loss, and a surface constraint, plus a new variance-scaled initialization.
Key results
- On ShapeNet, mean CD L2 x100 of 0.0075 and NC 0.9737, versus 0.0086 and 0.9723 for GP (prior best listed).
- On D-FAUST, CD L1 0.009 and F-Score 0.986 vs. 0.015 and 0.975 for GP.
- On 3DScene, CD L2 x100 0.094 vs. 0.10 for GP; ablation shows 3-step MSP with layers L4,L6,L8 gives CD L2 x100 0.036 vs 0.042 for linear layers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MultiPull instead of GP | surface reconstruction · ShapeNet | CD_L2 ×100: 0.0086 -> 0.0075 | Helped |
| Use MultiPull instead of GP | surface reconstruction · D-FAUST | CD_L1: 0.015 -> 0.009 | Helped |
| Use MultiPull instead of GP | surface reconstruction · Thingi10K | F-Score 0.01: 0.948 -> 0.953 | Helped |
| Use MultiPull instead of GP | surface reconstruction · 3DScene | NC: 0.903 -> 0.918 | Helped |
| Add frequency features instead of linear layers (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.042 -> 0.036 | Helped |
| Add frequency features instead of linear layers (ablation) | MultiPull · FAMOUS | NC: 0.920 -> 0.948 | Helped |
| Increase pulling steps from 1 to 3 (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.040 -> 0.036 | Helped |
| Increase pulling steps from 1 to 3 (ablation) | MultiPull · FAMOUS | NC: 0.926 -> 0.948 | Helped |
| Add distance-aware reconstruction loss to pulling loss (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.0443 -> 0.0383 | Helped |
| Add gradient consistency loss (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.0383 -> 0.0367 | Helped |
| Add surface constraint loss (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.0367 -> 0.0352 | Helped |
| Use the proposed initialization instead of random initialization (ablation) | MultiPull · FAMOUS | CD_L2 ×100: 0.042 -> 0.035 | Helped |
Surprising findings
- Adding multi-step pulling with a plain linear network (Linear+MSP) improves over Linear+Pull, and the paper reports that a plain PointMLP encoder with MSP also helps, suggesting the multi-step scheme is useful independent of frequency features.
Limitations
- Training time and computational efficiency are acknowledged as a limitation compared with NP.
- Multi-resolution feature methods such as NGLOD and Instant-NGP are not yet integrated.
- Gains over GP are small on several benchmarks (e.g. 3DScene CD L1 is tied with GP).
Experimental setup
Datasets: ShapeNet FAMOUS Surface Reconstruction Benchmark (SRB) Thingi10K D-FAUST 3DScene KITTI
Benchmarks: ShapeNet FAMOUS SRB Thingi10K D-FAUST 3DScene KITTI
Metrics: CD L1 CD L2 Normal Consistency F-Score
Baselines: ATLAS DSDF NP PCP GP IGR GenSDF FGC SAP BACON CAP P2M
Builds on: Neural-Pull Multiplication Filter Network FOCAL Gradient-based point cloud reconstruction (GP)
Closest papers by meaning
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.89)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.88)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.87)
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (NeurIPS 2024 · similarity 0.87)
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.87)
- High-Fidelity Lightweight Mesh Reconstruction from Point Clouds (CVPR 2025 · similarity 0.87)
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.