Pengyang Zhang, Wenlihan Lu, Shijian Gao
The Hong Kong University of Science and Technology (Guangzhou)
Paper · Dataset · Code · Citation
TeRFS extends radio-field synthesis from static scenes to spatio-temporal reconstruction. It couples an anisotropic spherical Gaussian (ASG) directional basis for sharp angular structure with per-lobe Gaussian temporal envelopes for path lifecycles. This design supports an explicit birth-and-death mechanism to capture multipath reorganization as moving scatterers create and block propagation paths.
In dynamic scenes, TeRFS represents the propagation environment, models how multipath components evolve over time, and synthesizes the angular spectrum at each receiver, capturing radio-field changes induced by moving scatterers. This provides a foundation for environment-aware wireless applications under high mobility.
Motivated by the sparse, directional structure of RF multipath, TeRFS adopts ASG to represent sharp angular peaks that spherical harmonics (SH) tend to smooth out.
Expressive capacity of the ASG basis, illustrated by fitting HFSS far-field patterns.
TeRFS builds on this directional expressiveness with per-lobe temporal envelopes, enabling explicit modeling of multipath birth and death.
| Evaluation | Key result |
|---|---|
| Temporal interpolation | 75% of test samples have absolute RSS error ≤ 3.26 dB |
| Spatial synthesis | 11.5% lower mean MSE than NeRF² |
| Training efficiency | 6.9× faster training than NeRF² |
| Representation efficiency | 63.1% fewer primitives than explicit baselines |
Evaluated on TeRFS-Dynamic using a single RTX 4090.
Reconstruction quality versus training time under default training schedules.
The TeRFS-Dynamic dataset captures radio-field evolution in a simulated outdoor campus, where multiple vehicles and a UAV create and block propagation paths.
| NeRF² | TeRFS-Dynamic | |
|---|---|---|
| Setting | Static scenes | Dynamic temporal sequence |
| Scatterers | Stationary | Moving vehicles and UAV |
| Scene | Indoor / semi-indoor | Outdoor campus |
Four timesteps showing multipath reorganization as vehicles pass near the roadside transmitter.
- Refining ASG representations for more accurate and efficient modeling of dynamic radio environments.
- Understanding multipath evolution through deeper insights into the underlying propagation mechanisms.
- Extending to diverse scenarios with varied environments and mobility patterns.
Code release coming soon.
If you find TeRFS or its dataset useful, please cite:
@article{zhang2026terfs,
title = {{TeRFS}: Temporal-Evolving Radio Field Synthesis},
author = {Zhang, Pengyang and Lu, Wenlihan and Gao, Shijian},
journal = {arXiv preprint arXiv:2605.02359},
year = {2026}
}Pengyang Zhang · zgotohdx@outlook.com