Research map / 3D vision / Gaussian splatting
NEAT: Distilling 3D Wireframes from Neural Attraction Fields
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 NEAT, a matching-free rendering-distillation method that learns 3D line segments via a neural attraction field built on VolSDF and distills sparse global 3D junctions via bipartite matching for multi-view 3D wireframe reconstruction.
Problem
Multi-view 3D wireframe reconstruction from posed images and 2D wireframe detections, where prior line-matching pipelines suffer from view-dependent endpoint errors and fragmentation.
Approach
Replace cross-view 2D line matching with volume rendering of 3D line segments from 2D attraction-field pixels using a neural attraction field, then distill a sparse set of learnable global 3D junctions from the noisy line cloud via DBSCAN pseudo-labels and Hungarian bipartite matching, jointly optimized from scratch.
Key results
- On DTU, NEAT achieves average ACC-J 0.7718 and ACC-L 0.8002 versus Line3D++@HAWPv3 ACC-J 0.9019 and ACC-L 0.8133 (lower is better).
- On BlendedMVS, NEAT average ACC-J 0.1949 and ACC-L 0.1802 versus Line3D++@HAWPv3 0.3743 and 0.3545.
- For 3D Gaussian Splatting, NEAT's 549 junctions serve as initialization instead of SfM points, reported as about 20 times fewer initial 3D points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NEAT instead of LiMAP | 3D wireframe reconstruction · DTU Avg. | ACC-J: 1.0944 -> 0.7718 | Helped |
| Use NEAT instead of Line3D++@HAWP | 3D wireframe reconstruction · BlendedMVS Avg. | ACC-L: 0.3545 -> 0.1802 | Helped |
| Add view direction to the NEAT field parameterization (ablation) | NEAT · DTU-24 | ACC-L: 0.847 -> 0.678 | Helped |
| Add view direction to the NEAT field parameterization (ablation) | NEAT · DTU-105 | ACC-J: 0.822 -> 0.749 | Helped |
| Add DBScan clustering during junction perceiving (ablation) | NEAT with view direction · DTU-24 | ACC-J: 0.796 -> 0.639 | Helped |
| Add DBScan clustering during junction perceiving (ablation) | NEAT with view direction · DTU-105 | ACC-J: 0.749 -> 0.720 | Helped |
| Add DBScan clustering during junction perceiving (ablation) | NEAT with view direction · DTU-24 | #Lines: 827 -> 860 | Hurt |
| Add DBScan clustering during junction perceiving (ablation) | NEAT with view direction · DTU-105 | #Lines: 557 -> 621 | Helped |
Surprising findings
- Viewing-direction parameterization reduces ACC errors but also reduces the number of junctions and lines, and clustering further improves accuracy, indicating a trade-off in completeness.
Limitations
- Volume rendering struggles with inside-out scenes such as ScanNet-like captures.
- Final wireframe quality depends on 2D wireframe detection quality (HAWPv3), e.g., motion blur failures.
- Predefined number of global junctions (e.g., 1024) limits scalability to large scenes.
Experimental setup
Datasets: DTU BlendedMVS ABC
Benchmarks: DTU BlendedMVS
Metrics: ACC-J ACC-L COMP-L #Lines #Junctions
Baselines: Line3D++ LiMAP LSD
Builds on: VolSDF HAWPv3 Line3D++ LiMAP 3D Gaussian Splatting
Closest papers by meaning
- Wired Perspectives: Multi-View Wire Art Embraces Generative AI (CVPR 2024 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.84)
- Structure from Collision (CVPR 2025 · similarity 0.84)
- GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction (NeurIPS 2024 · similarity 0.84)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.84)
- TurboSL: Dense Accurate and Fast 3D by Neural Inverse Structured Light (CVPR 2024 · similarity 0.84)
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.