Research map / 3D vision / NeRF and radiance fields
Can NeRFs "See" without Cameras?
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.
Redesigns NeRF to learn indoor floorplans from sparse WiFi RSSI measurements by modeling line-of-sight plus first-order reflections over a plausible voxel set with discrete orientations, trained in two stages.
Problem
Inferring the indoor floorplan of a home from sparse multipath WiFi signal-power measurements, where reflections are unknown and entangled with the line-of-sight signal.
Approach
EchoNeRF gives each voxel an opacity and a discrete orientation, models received power as LoS plus first-order reflections from a geometrically pruned plausible set of voxels, and uses a two-stage training schedule that first fits LoS and then uses LoS-derived gradients to supervise orientations and fit reflections.
Key results
- EchoNeRF achieves Wall_IoU 0.38 and F1 0.55 with 2000 receiver locations, versus 0.14 / 0.24 for NeRF2 and 0.27 / 0.42 for the LoS-only ablation.
- EchoNeRF RPE is 3.56 at 2000 receivers, better than NeRF2 (4.36) but worse than the MLP baseline (1.03).
- Robustness: floorplans remain legible down to about 30 dB SNR, and break down at 10 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add first-order reflections to the LoS-only model (ablation) | EchoNeRF · 2000 receiver locations | Wall_IoU: 0.27 ± 0.07 -> 0.38 ± 0.06 | Helped |
| Add first-order reflections to the LoS-only model (ablation) | EchoNeRF · 2000 receiver locations | F1 score: 0.42 ± 0.10 -> 0.55 ± 0.06 | Helped |
| Add first-order reflections to the LoS-only model (ablation) | EchoNeRF · 1000 receiver locations | Wall_IoU: 0.25 ± 0.04 -> 0.32 ± 0.06 | Helped |
| Add first-order reflections to the LoS-only model (ablation) | EchoNeRF · 1000 receiver locations | F1 score: 0.39 ± 0.06 -> 0.48 ± 0.05 | Helped |
| Use EchoNeRF instead of MLP for RSSI prediction | RSSI prediction · 2000 receiver locations | RPE: 1.03 -> 3.56 | Hurt |
| Use EchoNeRF instead of MLP for RSSI prediction | RSSI prediction · 1000 receiver locations | RPE: 0.65 -> 4.32 | Hurt |
| Increase material reflectivity (wood to concrete) | EchoNeRF · five materials across six scenes | Wall_IoU: 0.311 to 0.371 | Helped |
| Add RSSI measurement noise and lower SNR | EchoNeRF · SNR robustness across six floorplans | Wall_IoU: 0.371 -> 0.140 | Hurt |
Surprising findings
- A plain MLP achieves lower RSSI prediction error than NeRF2, suggesting NeRF2's implicit representation is not an advantage for RSSI interpolation.
- Vanilla L2 training on LoS plus reflection power failed to produce legible floorplans because LoS gradients dominate the reflection gradients.
Limitations
- Models only LoS and first-order reflections; higher-order reflections (about 5% of power) and furniture effects are ignored.
- Evaluated only in simulation (Sionna) on largely unfurnished ZInD floorplans; Sionna does not model through-wall penetration.
- Absolute Wall_IoU values are low (0.38 at best), so floorplans remain rough with blind spots and missing wall segments.
Experimental setup
Datasets: Zillow Indoor Dataset (ZInD) Sionna RT simulated 2.4 GHz WiFi RSSI
Benchmarks: ZInD simulated floorplans with Sionna RSSI
Metrics: Wall IoU F1 score RSSI prediction error (RPE)
Baselines: NeRF2 NeWRF Heatmap Segmentation MLP EchoNeRF_LoS
Builds on: NeRF NeRF2 NeWRF Friss equation
Closest papers by meaning
- NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds (CVPR 2023 · similarity 0.89)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.87)
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 2024 · similarity 0.87)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.86)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.85)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (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.