RadCloudSplat

Scatterer-Driven 3D Gaussian Splatting with Point-Cloud Priors for Radiomap Extrapolation

Yiheng Wang · Ye Xue · Shutao Zhang · Hongmiao Fan · Tsung-Hui Chang

Abstract

Radiomap represents the spatial distribution of wireless signal strength, critical for applications like network optimization and autonomous driving. However, constructing radiomap relies on measuring radio signal power across the entire system, which is costly in outdoor environments due to large network scales. We present RadSplatter, a framework that extends 3D Gaussian Splatting (3DGS) to radio frequencies for efficient and accurate radiomap extrapolation from sparse measurements. RadSplatter models environmental scatterers and radio paths using 3D Gaussians, capturing key factors of radio wave propagation. It employs a relaxed-mean (RM) scheme to reparameterize the positions of 3D Gaussians from noisy and dense 3D point clouds. A camera-free 3DGS-based projection is proposed to map 3D Gaussians onto 2D radio beam patterns. Furthermore, a regularized loss function and recursive fine-tuning using highly structured sparse measurements in real-world settings are applied to ensure robust generalization. Experiments on synthetic and real-world data show state-of-the-art extrapolation accuracy and execution speed.

Framework Overview

How It Works

☁️

LiDAR Point Cloud

Environment sensing

→

🎯

RM Selection

Key scatterers

→

📐

Camera-Free Proj

Beamspace mapping

→

⚡

EM Splatting

Complex attenuation

→

📊

RSS Radiomap

Beam-wise output

Step 1 — A trainable selection matrix T and bias B reparameterize scatterer positions from dense, noisy LiDAR point clouds, selecting N key virtual scatterers while remaining differentiable for end-to-end optimization.

Step 2 — Camera-free projection maps 3D Gaussians onto the 2D beamspace via learned MLPs: one recovers beamspace means, another recovers precision (inverse covariance), and a third encodes distance-dependent complex attenuation with phase from the base station.

Step 3 — Gaussians are depth-sorted by distance to the target location, then electromagnetically splatted: RSS-encoded spherical harmonics modulate the view-dependent signal, while complex-valued opacity (amplitude × e^j·phase) is alpha-blended with cumulative transmittance.

Step 4 — The final beam-wise RSS is the squared magnitude of the accumulated coherent signal, trained under a composite loss combining MSE, total variation regularization, and structural penalties on T and B for robust extrapolation.

Results

Synthetic Dataset (City A)

Method	MAE (dB) ↓	Inference Time (s)
Kriging	11.307	0.010
VAE	10.306	0.007
NeRF²	9.867	0.020
RadSplatter (ours)	7.564	0.004

Ablation Study

Method	MAE (dB) ↓
RadSplatter (full)	7.564
Random-Points-Opt	8.167
Random-Initial-Opt	10.463
Random-Initial-Fixed	53.234

Real-world Dataset (City B)

Method	MAE (dB) ↓	Inference Time (s)
Kriging	9.656	0.183
VAE	8.728	0.011
NeRF²	7.313	0.585
RadSplatter (ours)	7.035	0.018

Conclusion

In this work, we first extended 3DGS to the radio frequency domain, leveraging camera-free RadSplatter to extrapolate RSSs with high accuracy from sparse measurements in outdoor environments. By efficiently selecting the means of key virtual scatterers from dense point clouds aided by the RM parameterization model, the model captured intricate multi-path propagation characteristics. Experiments and analysis validated the effectiveness of these scatterers, advancing the state-of-the-art in wireless network modeling and extrapolation performance and highlighting the transformative potential of integrating advanced 3D modeling techniques with wireless propagation analysis for next-generation applications in the radio domain.

Citation

@article{wang2025radsplatter,
  title     = {RadSplatter: Extending 3D Gaussian Splatting to Radio
               Frequencies for Wireless Radiomap Extrapolation},
  author    = {Wang, Yiheng and Xue, Ye and Zhang, Shutao
               and Fan, Hongmiao and Chang, Tsung-Hui},
  journal   = {arXiv preprint arXiv:2502.12686},
  year      = {2025}
}

This work was accepted by IEEE INFOCOM 2026. We thank Dr. Xiaopeng Zhao and the authors of NeRF² for making their code and dataset available.