Research map / 3D vision / Face and shape
3D Neural Edge Reconstruction
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 EMAP, which learns a neural unsigned distance field (UDF) from multi-view edge maps with an unbiased rendering formulation and extracts parametric 3D lines and Bézier curves from it.
Problem
Reconstructing complete 3D edges (both straight lines and curves) from multi-view posed images, where prior line-mapping methods ignore curves and neural edge-field methods suffer rendering bias and fragile post-processing.
Approach
Model 3D edges as the zero-level neighborhood of a UDF trained by volume rendering of 2D edge maps with an unbiased monotonic density (from NeuS/NeuralUDF), then extract edge points by iterative shifting along inverse normals, edge directions via SVD on neighboring normals, and fit lines and Bézier curves via polyline connection and merging.
Key results
- On ABC-NEF with DexiNed edges, EMAP achieves completeness 8.9 mm versus 15.6 for LIMAP and 16.9 for NEF's best setting, and F5 of 59.1% versus 39.0% for LIMAP (LSD).
- On DTU, EMAP reports mean R5 of 77.5% and P5 of 75.4%, the highest recall and precision among compared methods.
- Using the trained UDF to initialize MonoSDF yields finer geometric details on Replica (qualitative only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EMAP with DexiNed instead of LIMAP with LSD | 3D edge reconstruction · ABC-NEF | F5: 39.0 -> 59.1 | Helped |
| Use EMAP instead of LIMAP | 3D edge reconstruction · DTU | R5: 74.9 -> 77.5 | Helped |
| Add query point shifting (ablation) | EMAP parametric edge extraction · ABC-NEF | R5: 29.2 -> 56.4 | Helped |
| Add Bézier curves (ablation) | EMAP parametric edge extraction · ABC-NEF | Comp: 12.1 -> 8.9 | Helped |
| Add edge merging (ablation) | EMAP parametric edge extraction · ABC-NEF | P5: 45.3 -> 62.9 | Helped |
| Add endpoint merging (ablation) | EMAP parametric edge extraction · ABC-NEF | F5: 54.0 -> 59.1 | Helped |
| Add edge refinement (ablation) | EMAP · dark scenes (qualitative) | edge artifact reduction | Helped |
| Initialize MonoSDF geometry MLPs with EMAP's pretrained UDF MLPs | MonoSDF · Replica | geometric detail reconstruction | Helped |
Surprising findings
- NEAT fails on ABC-NEF because it depends heavily on texture input.
- LIMAP achieves the best line accuracy and line-segment performance, slightly beating EMAP on lines while EMAP wins on curves.
Limitations
- Relies on quality of pretrained 2D edge detectors; noisy edges in dark scenes require a refinement step.
- Lower line accuracy than LIMAP on some metrics; NEAT and NEF struggle on textureless or larger scenes.
- Scene-level evaluation on Replica and Tanks & Temples is qualitative only due to lack of ground-truth edges.
Experimental setup
Datasets: ABC-NEF DTU Replica Tanks & Temples
Benchmarks: ABC-NEF DTU Replica Tanks & Temples
Metrics: Accuracy Completeness Recall Precision F-score Edge Direction Consistency
Baselines: NEF NEAT LIMAP
Builds on: NeRF NeuS NeuralUDF NEF LIMAP MonoSDF
Closest papers by meaning
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.88)
- Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes (CVPR 2024 · similarity 0.87)
- MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction (NeurIPS 2025 · similarity 0.86)
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.86)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.86)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 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.