Research map / 3D vision / Depth and stereo
NFL-BA: Near-Field Light Bundle Adjustment for SLAM in Dynamic Lighting
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 NFL-BA, a bundle adjustment loss that models dynamic near-field (co-located light) shading inside neural-rendering SLAM, improving tracking and mapping on C3VD endoscopy and self-captured indoor flashlight scenes.
Problem
SLAM systems assume static distant illumination and fail when the only light source is co-located with the moving camera, as in endoscopy and search-and-rescue, where near-field shading creates strong view-dependent artifacts that degrade tracking and mapping.
Approach
Replace the photometric bundle adjustment rendering loss with a near-field-lighting image formation model (albedo times inverse-square fall-off times Lambertian shading) integrated into the volumetric or Gaussian rendering equation, using surface normals from occupancy gradients or Gaussian shortest axes; it is a drop-in change to existing neural-rendering SLAM.
Key results
- On C3VD with PPSNet depth, NFL-BA reduces MonoGS tracking error from 3.48 to 2.18 mm (37%) and Chamfer error from 1.59 to 0.99 mm.
- On self-captured indoor scenes with MonoGS, NFL-BA reduces average translational ATE by roughly 35% across four scenes.
- NFL-BA does not improve every case: EndoGSLAM rotation/translation and NICE-SLAM results are mixed across sequences in the per-sequence tables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Photo-BA with NFL-BA | MonoGS with PPS-Net depth · C3VD | ATE_t (mm): 3.48 -> 2.18 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS with PPS-Net depth · C3VD | Chamfer (mm): 1.59 -> 0.99 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS with DA-Hybrid depth · C3VD | ATE_t (mm): 4.63 -> 2.35 | Helped |
| Replace Photo-BA with NFL-BA | EndoGSLAM with PPS-Net depth · C3VD | ATE_t (mm): 3.03 -> 2.62 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS · Guitars | ATE_t (m): 0.30 -> 0.18 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS · Porch | ATE_t (m): 0.50 -> 0.35 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS · Pool | ATE_t (m): 0.41 -> 0.30 | Helped |
| Replace Photo-BA with NFL-BA | MonoGS · Stairs | ATE_t (m): 0.36 -> 0.20 | Helped |
| Use NFL-BA instead of Photo-BA | MonoGS · C3VD oracle depth | ATE_r (°): 1.11 -> 1.49 | Hurt |
| Use NFL-BA instead of Photo-BA | EndoGSLAM · C3VD oracle depth | Chamfer (mm): 0.85 -> 0.97 | Hurt |
| Set angular attenuation coefficient β to non-zero values instead of β=0 (ablation) | MonoGS + NFL-BA · Guitars and Porch | ATE_T: 0.17±0.01; 0.33±0.16 -> β=0.25: 0.17±0.003; 0.25±0.013 | Mixed |
| Use β=0.50 instead of β=0 (ablation) | MonoGS + NFL-BA · Guitars | LPIPS: 0.36±0.02 -> 0.40±0.030 | Hurt |
Surprising findings
- Center-cropping the image improves MonoGS tracking nearly as much as NFL-BA but substantially degrades reconstruction density and Chamfer distance.
- Per-sequence tables show NFL-BA sometimes worsens tracking substantially (e.g., NICE-SLAM and MonoGS with DA-Hybrid on some sequences), indicating high variance.
Limitations
- Models only diffuse reflectance and direct illumination; specular highlights, subsurface scattering and inter-reflections are not handled, and a 0.9 intensity mask is used as a workaround.
- Angular attenuation coefficient beta is fixed to zero, justified empirically but not learned.
- Performance degrades on Colon10K sequences with more specular highlights.
Experimental setup
Datasets: C3VD Colon10K self-captured indoor scenes (iPhone 15 Pro)
Benchmarks: C3VD Colon10K self-captured indoor scenes
Metrics: ATE translation ATE rotation Chamfer distance LPIPS
Baselines: MonoGS (Photo-BA) EndoGSLAM (Photo-BA) NICE-SLAM (Photo-BA) center-crop MonoGS
Builds on: MonoGS EndoGSLAM NICE-SLAM Photometric bundle adjustment
Closest papers by meaning
- Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB-D Cameras (CVPR 2024 · similarity 0.83)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.81)
- 4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians (CVPR 2025 · similarity 0.81)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.81)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.81)
- IBD-SLAM: Learning Image-Based Depth Fusion for Generalizable SLAM (CVPR 2024 · similarity 0.81)
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.