Research map / 3D vision / Face and shape
GeRaF: Neural Geometry Reconstruction from Radio Frequency Signals
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.
Introduces GeRaF, a neural implicit SDF method with physics-based RF volumetric rendering and lensless sampling/alpha blending that reconstructs millimeter-level object geometry from real 77 GHz mmWave radar scans.
Problem
Near-range 3D geometry reconstruction from low-resolution, noisy, lensless radio frequency (mmWave) signals is hard because volumetric rendering over full 3D space is computationally intractable and RF reflections are specular rather than diffuse.
Approach
Model RF reflection with a shifted Lambertian specular term inside SDF-based volumetric rendering, share opacity and transmittance across antennas via lensless sampling along a primary direction, reuse per-antenna signal tracing through a memory bank, and mask unreliable low-power samples during training.
Key results
- Outperforms the matched filter baseline on all 10 objects, e.g. Bottle F1 0.97 vs 0.44 and Chamfer distance 0.05 mm vs 0.35 mm.
- On occluded objects, Knife* F1 improves from 0.45 to 0.83 and Chamfer distance drops from 0.84 mm to 0.26 mm.
- Novel view synthesis on the Bunny achieves PSNR around 30 dB on normalized matched filter images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeRaF instead of the Matched Filter baseline | 3D reconstruction · Unoccluded objects (Wrench, Bunny, Elephant, Bottle, Knife, Star) | F1: 0.52 -> 0.88; 0.56 -> 0.81; 0.65 -> 0.79; 0.44 -> 0.97; 0.44 -> 0.71; 0.80 -> 0.86 | Helped |
| Use GeRaF instead of the Matched Filter baseline | 3D reconstruction · Unoccluded objects (Wrench, Bunny, Elephant, Bottle, Knife, Star) | CD (mm): 0.29 -> 0.07; 0.03 -> 0.01; 0.25 -> 0.14; 0.35 -> 0.05; 0.59 -> 0.49; 0.60 -> 0.50 | Helped |
| Use GeRaF instead of the Matched Filter baseline | 3D reconstruction · Occluded objects (Knife, Elephant, Boat, Wrench) | F1: 0.45 -> 0.83; 0.42 -> 0.52; 0.51 -> 0.69; 0.63 -> 0.86 | Helped |
| Use GeRaF instead of the Matched Filter baseline | 3D reconstruction · Occluded objects (Knife, Elephant, Boat, Wrench) | CD (mm): 0.84 -> 0.26; 0.52 -> 0.46; 0.35 -> 0.30; 0.29 -> 0.14 | Helped |
| Increase the number of radar scanning planes (ablation) | GeRaF · Bunny reconstruction | Reconstruction quality | Helped |
| Vary the number of temporal samples (ablation) | GeRaF · Geometry reconstruction | Reconstruction performance | No clear effect |
| Increase positional encoding frequency levels (ablation) | GeRaF · Bunny reconstruction | Fine-detail reconstruction | Helped |
| Increase the number of sampling rays (ablation) | GeRaF · Geometry reconstruction | Reconstruction performance | Hurt |
| Reduce the number of depth samples (ablation) | GeRaF · Geometry reconstruction | Reconstruction fidelity | Hurt |
| Add dynamic loss masking (ablation) | GeRaF · Bunny reconstruction | Geometric fidelity | Helped |
| Use the Lambertian directional term instead of removing it (ablation) | GeRaF · Bunny reconstruction | Chamfer Distance | Helped |
Surprising findings
- Denser ray sampling gave worse reconstruction despite higher cost, attributed to RF noise slowing convergence.
- Removing the Lambertian directional term changed visual output only minimally, though quantitative error rose.
Limitations
- Surfaces never reflecting back to the radar in any scan cannot be reconstructed.
- Only metallic objects tested; the simplified Lambertian RF reflection model is insufficient for other materials and accurate forward rendering.
- Thick occlusions and multi-level surfaces are not handled; computational cost remains much higher than vision rendering.
Experimental setup
Datasets: self-collected 77 GHz mmWave radar scans of real objects (TI 1843BOOST with DCA1000EVM on Franka Research 3) Scaniverse iPhone LiDAR/camera scans as ground truth geometry
Benchmarks: self-collected real mmWave radar dataset
Metrics: F1-score (tau=0.01) Chamfer distance (mm) PSNR (dB) for novel view synthesis
Baselines: Matched filter (MF) summation with Poisson surface reconstruction
Builds on: NeRF NeuS matched filter Lambertian reflection model positional encoding
Closest papers by meaning
- RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis (NeurIPS 2026 · similarity 0.86)
- Radio Frequency Ray Tracing with Neural Object Representation for Enhanced RF Modeling (CVPR 2025 · similarity 0.84)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.84)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.83)
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors (ICLR 2024 · similarity 0.83)
- NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields (ICLR 2025 · similarity 0.82)
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.