Research map / 3D vision / 3D generation
Swift-Mapping: Online Neural Implicit Dense Mapping in Urban Scenes
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 Swift-Mapping, an online neural implicit dense mapping framework for urban scenes built on a sparse octree feature structure (neural implicit octomap) with hierarchical latent vectors, instead of keyframes.
Problem
Online dense mapping of large, dynamic urban scenes with neural implicit representations is hard because dense voxel grids and keyframe-based optimization are too memory-hungry and slow, and prior methods are limited to indoor scenes.
Approach
A neural implicit octomap (NIO) stores sparse octree-organized latent features with distance-adaptive voxel resolution, and hierarchical latent vectors fused by max pooling with learnable memorization parameters replace keyframe lists; sparse hybrid sampling of RGB and LiDAR-completed depth supports noisy urban depth.
Key results
- On Replica, Swift-Mapping reaches Acc 1.6219 and Comp Ratio 84.42 versus NICE-SLAM's 1.7510 and 86.09, at 0.71 FPS versus 0.45 FPS.
- On VKITTI2 rendering, L2 loss 307.10 and PSNR 23.12 versus NICE-SLAM's 930.32 and 19.15.
- The paper claims more than 10x faster reconstruction than offline baselines, which take 10 minutes of training on urban scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Swift-Mapping with the NIO representation and hierarchical latent vectors instead of NICE-SLAM | Online dense mapping · ScanNet | PSNR: 22.81 -> 24.23 | Helped |
| use Swift-Mapping with the NIO representation and hierarchical latent vectors instead of NICE-SLAM | Online dense mapping · Replica | L2 Loss: 179.51 -> 88.32 | Helped |
| use Swift-Mapping with the NIO representation and hierarchical latent vectors instead of Instant-NGP | Urban-scene rendering · VKITTI2 | PSNR: 20.87 -> 23.12 | Helped |
| use Swift-Mapping with the NIO representation and hierarchical latent vectors instead of Instant-NGP | Urban-scene rendering · KITTI | L2 Loss: 516.54 -> 359.68 | Helped |
| use Swift-Mapping instead of NICE-SLAM | Replica reconstruction · Replica | FPS: 0.45 -> 0.71 | Helped |
| increase octree depth while keeping the smallest voxel resolution fixed (ablation) | Swift-Mapping · VKITTI2 | L2 Loss: 712.09 -> 535.32 -> 420.03 -> 339.68 -> 317.01 | Helped |
| increase octree depth while keeping the smallest voxel resolution fixed (ablation) | Swift-Mapping · VKITTI2 | FPS: 0.46 -> 0.39 -> 0.36 -> 0.30 -> 0.26 | Hurt |
| add trainable memorization parameters αk and βk (ablation) | Swift-Mapping · ScanNet | PSNRmem: 16.52 -> 21.37 | Helped |
| add trainable memorization parameters αk and βk (ablation) | Swift-Mapping · VKITTI2 | PSNRmem: 17.83 -> 23.29 | Helped |
| add trainable memorization parameters αk and βk (ablation) | Swift-Mapping · ScanNet | PSNRtrain: 23.75 -> 23.76 | No clear effect |
Surprising findings
- On Replica, the method is only comparable to NICE-SLAM, which is designed for indoor scenes, while being stronger in urban scenes.
- Under fixed iteration counts, NICE-SLAM's PSNR inference on VKITTI2 drops sharply compared to its training PSNR, suggesting forgetting.
Limitations
- Indoor performance is only comparable to NICE-SLAM rather than superior.
- Urban evaluation uses only 2D metrics because ground-truth meshes are unavailable.
- Sky regions are masked out during rendering because depth is unknown there.
Experimental setup
Datasets: Replica ScanNet KITTI VKITTI2 nuScenes
Benchmarks: Replica ScanNet KITTI VKITTI2 nuScenes
Metrics: PSNR MSE L2 loss Accuracy Completion ratio FPS
Baselines: iMAP NICE-SLAM Instant-NGP Mip-NeRF 360
Builds on: NeRF NICE-SLAM iMAP ConvOnet CompletionFormer
Closest papers by meaning
- ViiNeuS: Volumetric Initialization for Implicit Neural Surface Reconstruction of Urban Scenes with Limited Image Overlap (CVPR 2025 · similarity 0.88)
- Dynamic 3D Gaussian Fields for Urban Areas (NeurIPS 2024 · similarity 0.87)
- 3D LiDAR Mapping in Dynamic Environments using a 4D Implicit Neural Representation (CVPR 2024 · similarity 0.87)
- Multi-Level Neural Scene Graphs for Dynamic Urban Environments (CVPR 2024 · similarity 0.85)
- Benchmarking Implicit Neural Representation and Geometric Rendering in Real-Time RGB-D SLAM (CVPR 2024 · similarity 0.85)
- Implicit Zoo: A Large-Scale Dataset of Neural Implicit Functions for 2D Images and 3D Scenes (NeurIPS 2024 · similarity 0.85)
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.