Research map / Point clouds / Shape and surface completion
3D-SLNR: A Super Lightweight Neural Representation for Large-scale 3D Mapping
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 3D-SLNR, a feature-free neural SDF representation built from band-limited local SDFs anchored at support points, with a parallel CUDA local-SDF detection algorithm and prune-and-expand densification, achieving state-of-the-art large-scale mapping at under 1/5 the memory.
Problem
Large-scale 3D mapping from point clouds requires expressive scene representations, but existing methods are either memory-hungry sparse grids or rely on high-dimensional latent features.
Approach
A global SDF is defined as a weighted average of local SDFs, each a tiny MLP basis function in a learnable local frame (position, rotation, scale of each support point), with no latent features; a parallel grid-index algorithm detects overlapping local SDFs per sample in real time, and a prune-and-expand strategy adapts the support points.
Key results
- On Matterport3D, F-score 94.4 with 13.7 MB map memory versus SHINE-Mapping 92.8 at 167.4 MB.
- On Newer College Quad, F-score 92.7 with 3.3 MB versus PIN-SLAM 88.3 at 18.8 MB.
- Ablation shows adding rotation, scale and position raises F-score from 82.5 to 85.2 on Math Institute, and prune-and-expand reduces memory from 9.9 to 8.0 MB while improving F-score to 85.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 3D-SLNR instead of NKSR | 3D mapping on Matterport3D · Matterport3D | F-score: 94.2 -> 94.4 | Helped |
| Use 3D-SLNR instead of NKSR | 3D mapping on Matterport3D · Matterport3D | MC: 128.9 -> 13.7 | Helped |
| Use 3D-SLNR instead of SHINE-Mapping | 3D mapping on Newer College Quad · Newer College Quad | F-score: 91.2 -> 92.7 | Helped |
| Use 3D-SLNR instead of SHINE-Mapping | 3D mapping on Newer College Quad · Newer College Quad | MC: 15.8 -> 3.3 | Helped |
| Use 3D-SLNR instead of SHINE-Mapping | 3D mapping on Newer College Math Institute · Newer College Math Institute | F-score: 84.0 -> 85.3 | Helped |
| Use 3D-SLNR instead of SHINE-Mapping | 3D mapping on Newer College Math Institute · Newer College Math Institute | MC: 33.8 -> 8.0 | Helped |
| Use 3D-SLNR instead of PIN-SLAM | 3D mapping on street-level dataset · Street-level dataset street scene | F-score: 87.1 -> 88.6 | Helped |
| Add rotation parameters (ablation) | 3D-SLNR on Newer College Math Institute · Newer College Math Institute | F-score: 82.5 -> 84.7 | Helped |
| Add scaling parameters (ablation) | 3D-SLNR on Newer College Math Institute · Newer College Math Institute | F-score: 84.7 -> 85.0 | Helped |
| Add position parameters (ablation) | 3D-SLNR on Newer College Math Institute · Newer College Math Institute | F-score: 85.0 -> 85.2 | Helped |
| Add prune-and-expand (ablation) | 3D-SLNR on Newer College Math Institute · Newer College Math Institute | F-score: 85.2 -> 85.3 | Helped |
| Add prune-and-expand (ablation) | 3D-SLNR on Newer College Math Institute · Newer College Math Institute | MC: 9.9 -> 8.0 | Helped |
Surprising findings
- Prune-and-expand reduces memory because more support points are removed than added, unlike 3DGS densification which grows the primitive count.
Limitations
- Running efficiency is not yet optimized, as the authors note.
- Not yet integrated into an online SLAM system.
- Performance is sensitive to support point initialization density (overfitting or underfitting at extremes).
Experimental setup
Datasets: Matterport3D Newer College KITTI-360 self-collected street-level dataset
Benchmarks: Matterport3D Newer College KITTI-360
Metrics: Accuracy Completeness Precision Recall F-score memory consumption (MC)
Baselines: VDB-Fusion SPSR NKSR SHINE-Mapping PIN-SLAM
Builds on: Neural-implicit hash grid mapping (SHINE-Mapping) 3D Gaussian splatting densification Poisson surface reconstruction Point-NeRF
Closest papers by meaning
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.87)
- Neural Processing of Tri-Plane Hybrid Neural Fields (ICLR 2024 · similarity 0.86)
- Sharpening Neural Implicit Functions with Frequency Consolidation Priors (AAAI 2025 · similarity 0.85)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.84)
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.84)
- SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration (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.