Research map / Point clouds / Shape and surface completion
A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions
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.
Trains a lightweight attention-based UDF network once on a synthetic dataset of local point-cloud patches built from quadratic (smooth) and parametric sharp-feature functions, then reconstructs surfaces from unseen point clouds and can initialize unsupervised UDF methods.
Problem
Learning unsigned distance fields (UDFs) from point clouds usually needs slow per-shape unsupervised optimization or large shape-dataset supervised training that is costly and dataset-dependent.
Approach
Learn UDFs from local point-cloud patches synthesized from differential-geometry quadratic surfaces and simple sharp-feature functions (creases, cusps, corners, v-saddles), so training is shape-independent; a cross-attention module fuses patch features with query-relative vectors, plus a denoising displacement module.
Key results
- On ShapeNetCars with clean inputs, the method reports CD 0.611 vs 0.832 for GeoUDF and F1@0.005 0.645 (best on CD and F1@0.01 0.927).
- With 10% outliers on DeepFashion3D, CD 0.559 vs 2.558 for DUDF, with NC 0.909.
- Integrating LoSF as initialization reduces CAP-UDF runtime on SRB from 15.87 to 6.84 minutes and LevelSetUDF from 15.08 to 6.13 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LoSF-UDF instead of DUDF | UDF surface reconstruction · ShapeNetCars, clean | L1-Chamfer distance (×100): 0.773 -> 0.611 | Helped |
| Use LoSF-UDF instead of DUDF | UDF surface reconstruction · ShapeNetCars, clean | F-score (0.005): 0.697 -> 0.645 | Hurt |
| Use LoSF-UDF instead of LevelSetUDF | UDF surface reconstruction · ShapeNetCars, noise | L1-Chamfer distance (×100): 1.176 -> 0.614 | Helped |
| Use LoSF-UDF instead of LevelSetUDF | UDF surface reconstruction · ShapeNetCars, noise | F-score (0.005): 0.299 -> 0.546 | Helped |
| Use LoSF-UDF instead of GeoUDF | UDF surface reconstruction · ShapeNetCars, outliers | L1-Chamfer distance (×100): 2.598 -> 0.705 | Helped |
| Use LoSF-UDF instead of LevelSetUDF | UDF surface reconstruction · ShapeNetCars, outliers | F-score (0.005): 0.200 -> 0.437 | Helped |
| Use LoSF-UDF instead of DUDF | UDF surface reconstruction · DeepFashion3D, clean | L1-Chamfer distance (×100): 0.506 -> 0.548 | Hurt |
| Use LoSF-UDF instead of DUDF | UDF surface reconstruction · DeepFashion3D, clean | F-score (0.005): 0.586 -> 0.565 | Hurt |
| Use LoSF-UDF instead of LevelSetUDF | UDF surface reconstruction · DeepFashion3D, noise | L1-Chamfer distance (×100): 0.723 -> 0.614 | Helped |
| Use LoSF-UDF instead of GeoUDF | UDF surface reconstruction · DeepFashion3D, noise | F-score (0.005): 0.506 -> 0.541 | Helped |
| Use LoSF-UDF instead of DUDF | UDF surface reconstruction · DeepFashion3D, outliers | L1-Chamfer distance (×100): 2.558 -> 0.559 | Helped |
| Use LoSF-UDF instead of LevelSetUDF | UDF surface reconstruction · DeepFashion3D, outliers | F-score (0.005): 0.278 -> 0.463 | Helped |
Surprising findings
- A model trained only on synthetic local patches generalizes to real scans and to ShapeNet, DeepFashion3D and ScanNet without shape-specific training.
Limitations
- Cannot recover missing parts in real scans because it relies on local geometry independent of global shape.
- Needs sufficient point density (about 30 points per unit area) and a tuned patch radius.
Experimental setup
Datasets: ShapeNetCars DeepFashion3D ScanNet Threescan synthetic local shape-function patches
Benchmarks: ShapeNetCars DeepFashion3D ScanNet Threescan real range scan dataset
Metrics: L1 Chamfer distance F1-score normal consistency runtime parameter count
Baselines: CAP-UDF LevelSetUDF DUDF GeoUDF
Builds on: GeoUDF DCUDF PointNet CAP-UDF LevelSetUDF
Closest papers by meaning
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.91)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.89)
- MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction (NeurIPS 2025 · similarity 0.88)
- Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes (CVPR 2024 · similarity 0.88)
- Small Steps and Level Sets: Fitting Neural Surface Models with Point Guidance (CVPR 2024 · similarity 0.88)
- 2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight Model Reconstruction from Multi-view Images (CVPR 2024 · similarity 0.88)
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.