Research map / 3D vision / NeRF and radiance fields
2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight Model Reconstruction from Multi-view Images
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 2S-UDF, a two-stage UDF learning method for multi-view reconstruction of non-watertight models that first learns a coarse UDF with a bell-shaped density and then refines it by directly learning an unbiased, occlusion-aware weight function with ray truncation.
Problem
Learning unsigned distance fields from multi-view images via volume rendering is hard because converting UDF values to density must yield an unbiased, occlusion-aware, bounded weight function, so existing methods produce choppy or closed UDFs and fail on non-watertight models.
Approach
Decouple density and weight across two stages: Stage 1 uses an easily trainable bounded bell-shaped density for coarse UDF learning; Stage 2 truncates rays at the first surface crossing found from the Stage 1 UDF and directly learns a sigmoid-derivative weight function, bypassing density boundedness issues.
Key results
- On DeepFashion3D (DF3D#Ours), 2S-UDF reaches a mean Chamfer distance of 5.02e-3, versus 7.36e-3 for NeUDF and 8.53e-3 for NeuralUDF.
- On DF3D#NeuralUDF, 2S-UDF achieves mean Chamfer distance 1.96e-3, the best among compared methods (NeuralUDF 2.34e-3, NeUDF 2.26e-3).
- On DTU, 2S-UDF attains mean Chamfer distance 0.71e-3 versus 0.69e-3 for NeuralUDF, roughly matching it, and improves NeUDF's reconstruction by around 10%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use two-stage UDF learning instead of NeUDF | 2S-UDF · DF3D#Ours | Chamfer distance: 7.36 -> 5.02 | Helped |
| use two-stage UDF learning instead of NeUDF | 2S-UDF · DF3D#NeuralUDF | Chamfer distance: 2.26 -> 1.96 | Helped |
| refine NeUDF with Stage 2 weight adjustment | NeUDF initialization on DTU · DTU | Chamfer distance: 0.59 -> 0.41; 0.62 -> 0.51 | Mixed |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D #1 | Chamfer distance: 7.22 -> 4.55 | Helped |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D #7 | Chamfer distance: 2.46 -> 2.88 | Hurt |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D #8 | Chamfer distance: 3.38 -> 3.21 | Helped |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D LS-D0 | Chamfer distance: 6.04 -> 2.46 | Helped |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D NS-D1 | Chamfer distance: 1.46 -> 1.47 | No clear effect |
| add Stage 2 refinement after Stage 1 (ablation) | 2S-UDF · DeepFashion3D LS-C1 | Chamfer distance: 6.23 -> 2.14 | Helped |
| add Stage 1 coarse learning before Stage 2 (ablation) | 2S-UDF · DeepFashion3D #1 | Chamfer distance: 5.75 -> 4.55 | Helped |
| add Stage 1 coarse learning before Stage 2 (ablation) | 2S-UDF · DeepFashion3D #7 | Chamfer distance: 4.00 -> 2.88 | Helped |
| add Stage 1 coarse learning before Stage 2 (ablation) | 2S-UDF · DeepFashion3D #8 | Chamfer distance: 5.96 -> 3.21 | Helped |
Surprising findings
- Training with only Stage 2 from scratch yields reasonable but generally worse results than the two-stage pipeline, and the NeuralUDF two-stage variant produces fragmented output after Stage 2.
Limitations
- Cannot model transparent planes because light is cut off after the first surface.
- For watertight models, UDF learning is more complicated and SDF methods are advised.
- Mesh extraction with MeshUDF produces holes and staircase artifacts.
Experimental setup
Datasets: DeepFashion3D DTU BlendedMVS
Benchmarks: DeepFashion3D DTU
Metrics: Chamfer distance
Baselines: NeuS VolSDF NeAT NeuralUDF NeUDF
Builds on: NeuS NeuralUDF NeUDF HF-NeuS VolSDF
Closest papers by meaning
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.88)
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.88)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.87)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.87)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.86)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · 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.