Skip to content

RF-GS: Radio-Frequency Gaussian Splatting for Dynamic Electromagnetic Scene Representation

Radio-Frequency Gaussian Splatting for Dynamic Electromagnetic Scene Representation Rev 5Download

This directory contains the complete LaTeX source for the paper “RF-GS: Radio-Frequency Gaussian Splatting for Dynamic Electromagnetic Scene Representation” – a CVPR 2026/SIGGRAPH 2026 submission.

bgilbert1984/RF-GS-Radio-Frequency-Gaussian-Splatting-for-Dynamic-Electromagnetic-Scene-Representation: 3D Gaussian Splatting representation learned directly from raw radiofrequency measurements (Wi-Fi CSI, mmWave, UWB, RF-tomography) without any RGB or depth supervision.

Files

  • RF_GS_CVPR2026_Paper.tex – Main paper LaTeX source
  • references.bib – Bibliography file with all necessary citations
  • figures/ – Directory for paper figures (to be created)
  • supplementary/ – Directory for supplementary material

Paper Overview

This paper introduces the first 3D Gaussian Splatting approach for radio-frequency sensing, enabling real-time, high-fidelity reconstruction of dynamic scenes using only RF measurements (Wi-Fi CSI, mmWave, UWB, etc.).

Key Contributions

  1. RF-Native Supervision: Direct supervision of 3D Gaussians using complex-valued CSI and RF features
  2. Adaptive RF Density Control: Novel densification/pruning strategies for electromagnetic fields
  3. Real-time Rendering: 200+ fps GPU-optimized renderer for RF scenes

Results Summary

  • 9-14 dB PSNR improvement over RF-NeRF baselines
  • 35× faster training (14 min vs 8+ hours)
  • 200× faster rendering (214 fps vs 1 fps)
  • Real-world deployment with commodity Wi-Fi hardware

Figure Requirements

The paper requires the following figures to be generated:

Main Figures

  1. figures/teaser.pdf – Side-by-side RGB-GS vs RF-GS reconstruction
  2. figures/qualitative.pdf – Qualitative comparison (RF-NeRF, RF-InstantNGP, RF-GS, GT)
  3. figures/realworld_deployment.pdf – Real-world Wi-Fi setup and results
  4. figures/temporal_analysis.pdf – Temporal coherence analysis graph

Supporting Figures

  • Method diagrams showing RF-GS pipeline
  • Ablation study visualizations
  • Cross-modal performance comparisons
  • Gaussian density visualizations

Code Integration

This paper directly corresponds to the implementation in:

  • code/neural-gaussian-splats.py – Main RF-GS model
  • code/neural-correspondence.py – Supporting correspondence fields

Key classes referenced:

  • GaussianSplatModel – Core RF-GS representation
  • GaussianPointRenderer – Real-time rendering engine
  • Adaptive density control methods (prune(), densify(), fit_to_rf_data())

Compilation

To compile the paper:

pdflatex RF_GS_CVPR2026_Paper.tex
bibtex RF_GS_CVPR2026_Paper
pdflatex RF_GS_CVPR2026_Paper.tex
pdflatex RF_GS_CVPR2026_Paper.tex

Or use your preferred LaTeX editor (Overleaf recommended for collaboration).

Submission Timeline

  • Target Venue: CVPR 2026 (Deadline: November 2025)
  • Alternative: SIGGRAPH 2026 (Deadline: January 2026)
  • Status: Ready for figure generation and experimental validation

Impact Potential

This paper represents breakthrough work at the intersection of:

  • 3D Computer Vision (Gaussian Splatting)
  • Radio Frequency Sensing
  • Real-time Rendering
  • Privacy-Preserving Perception

Expected impact: High citation potential, strong venue acceptance probability, foundational work for RF-based 3D reconstruction.

Related Papers from This Codebase

This work is part of a series extractable from the neural-gaussian-splats.py codebase:

  1. RF-GS (this paper) – Core RF Gaussian Splatting
  2. Temporal Gaussian Splatting via Neural Correspondence Fields – 4D extension
  3. DOMA: Dynamic Object Motion Analysis in RF – Object tracking application
  4. Adaptive Density Control for Non-Optical Gaussian Splatting – Method generalization

Contact

Ben Gilbert – Text and Call me at 832-654-9435 | https://172-234-197-23.ip.linodeusercontent.com/?page_id=14