Research map / 3D vision / Depth and stereo
SNI-SLAM: Semantic Neural Implicit SLAM
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 SNI-SLAM, an RGB-D neural implicit SLAM system that jointly performs semantic mapping, surface reconstruction, and camera tracking using hierarchical semantic features and cross-attention fusion of appearance, geometry, and semantic features.
Problem
NeRF-based SLAM mostly builds RGB maps unsuitable for downstream tasks, while existing semantic NeRF methods require hours of offline training and do not support real-time semantic SLAM.
Approach
Coarse-to-fine hierarchical semantic feature planes, cross-attention fusion of geometry, appearance and semantic features, a one-way decoder design, and a feature-level loss that supervises the scene representation directly.
Key results
- On Replica (8 scenes), SNI-SLAM reaches ATE RMSE 0.456 cm vs 0.678 cm for ESLAM and 1.059 cm for Co-SLAM, and Depth L1 0.766 cm.
- On four Replica scenes, semantic mIoU averages 87.41% vs 82.37% for NIDS-SLAM.
- On ScanNet, average tracking RMSE is 6.54 cm vs 7.27 cm for ESLAM; on TUM, average RMSE is 4.35 cm vs 8.92 cm for Point-SLAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add coarse-to-fine hierarchical semantic mapping (ablation) | SNI-SLAM · Replica office0 | Tracking RMSE [cm]: 0.83 -> 0.55 | Helped |
| add coarse-to-fine hierarchical semantic mapping (ablation) | SNI-SLAM · Replica office0 | mIoU (%): 71.5 -> 84.1 | Helped |
| add feature loss (ablation) | SNI-SLAM with hierarchical semantic mapping · Replica office0 | Tracking RMSE [cm]: 0.55 -> 0.47 | Helped |
| add feature loss (ablation) | SNI-SLAM with hierarchical semantic mapping · Replica office0 | mIoU (%): 84.1 -> 85.0 | Helped |
| add fusion-based decoder design (ablation) | SNI-SLAM with hierarchical semantic mapping and feature loss · Replica office0 | Tracking RMSE [cm]: 0.47 -> 0.43 | Helped |
| add fusion-based decoder design (ablation) | SNI-SLAM with hierarchical semantic mapping and feature loss · Replica office0 | mIoU (%): 85.0 -> 85.3 | Helped |
| add cross-attention-based feature fusion (ablation) | SNI-SLAM with hierarchical semantic mapping, feature loss, and decoder design · Replica office0 | Tracking RMSE [cm]: 0.43 -> 0.33 | Helped |
| add cross-attention-based feature fusion (ablation) | SNI-SLAM with hierarchical semantic mapping, feature loss, and decoder design · Replica office0 | mIoU (%): 85.3 -> 86.0 | Helped |
Limitations
- Semantic evaluation is limited to four Replica scenes because NIDS-SLAM results exist only for those scenes.
- TUM RGB-D has no semantic ground truth, so semantic labels there come from DEVA segmentation rather than annotations.
Experimental setup
Datasets: Replica ScanNet TUM RGB-D
Benchmarks: Replica ScanNet TUM RGB-D
Metrics: Depth L1 Accuracy Completion Completion ratio ATE RMSE ATE mean mIoU FPS
Baselines: iMAP NICE-SLAM Vox-Fusion Co-SLAM ESLAM NIDS-SLAM Point-SLAM
Builds on: NeRF ESLAM NICE-SLAM StyleSDF Co-SLAM
Closest papers by meaning
- NICE-SLAM: Neural Implicit Scalable Encoding for SLAM (CVPR 2022 · similarity 0.89)
- Query Quantized Neural SLAM (AAAI 2025 · similarity 0.87)
- IBD-SLAM: Learning Image-Based Depth Fusion for Generalizable SLAM (CVPR 2024 · similarity 0.86)
- Benchmarking Implicit Neural Representation and Geometric Rendering in Real-Time RGB-D SLAM (CVPR 2024 · similarity 0.85)
- Swift-Mapping: Online Neural Implicit Dense Mapping in Urban Scenes (AAAI 2024 · similarity 0.84)
- MNE-SLAM: Multi-Agent Neural SLAM for Mobile Robots (CVPR 2025 · 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.