PC-TGS

Official implementation of:

Point-Cloud-Assistant Localized Statistical Channel Prediction by Tangent Gaussian Splatting

IEEE Transactions on Wireless Communications, 2026

Authors: Ye Xue, Yiheng Wang, Xinhua Shao, Qi Yan, Shutao Zhang, and Tsung-Hui Chang

PC-TGS is a point-cloud-assisted radio-frequency Gaussian Splatting framework for localized statistical channel prediction. It extrapolates the channel angular power spectrum (APS) to unmeasured outdoor locations using sparse RSRP measurements and dense LiDAR point clouds.

PC-TGS may also be searched as PCTGS or PCT-GS; the official method name used in the paper is PC-TGS.

• Paper: arXiv
• Published version: IEEE Xplore
• Project page: Project Page
• Code: This repository
• Citation: BibTeX
• License: MIT

Naming Note

The official method name is PC-TGS. Some source files retain the historical radsplatter_* filenames for implementation compatibility. In this repository, these files implement the PC-TGS framework and should not be interpreted as the official implementation of a separate paper.

---

Quick Start

Environment

# Python >= 3.8, CUDA-enabled GPU
pip install numpy scipy h5py pyyaml tqdm einops matplotlib
pip install torch torchvision torchaudio  # match your CUDA version

Note: The training dataset is proprietary. To use this code with your own data, prepare files matching the format expected by datasets_aps_new.py.

Training

python train_radsplatter_new.py \
  --config ./radsplatter_setting_new.yml \
  --gpu 0 \
  --mode train \
  --num_scatters 2000 \
  --world_size 1 \
  --num_max_angles 800 \
  --sh_up_iter 500

Evaluation

python train_radsplatter_new.py \
  --config ./radsplatter_setting_new.yml \
  --gpu 0 \
  --mode test \
  --num_scatters 2000 \
  --world_size 1 \
  --num_max_angles 800

---

Repository Structure

├── README.md
├── radsplatter_setting_new.yml         # Training and optimizer configuration
├── train_radsplatter_new.py            # Main training/testing entrance
├── radsplatter_model.py                # PC-TGS model (RM, SH coefficients, scatterer attributes)
├── radsplatter_render.py               # Tangent-plane projection + electromagnetic splatting
├── datasets_aps_new.py                 # Dataset loaders for RSRP and APS data
├── projection_utils.py                 # 3D-to-angular projection + Jacobian computation
├── complex_sh_utils_new.py             # Complex spherical harmonic evaluation
├── sh_utils.py                         # Real spherical harmonic utilities
├── pdf_utils.py                        # Gaussian PDF computation
├── prune_utils.py                      # Mahalanobis-based Gaussian filtering
├── loss_utils.py                       # Loss functions (L1, L2, SSIM, SmoothL1)
├── data_painter.py                     # APS visualization and data processing
└── utils.py                            # General tensor and rotation utilities

---

Citation

If you find this work helpful, please cite:

@article{xue2026point,
  title={Point-Cloud-Assistant Localized Statistical Channel Prediction by Tangent Gaussian Splatting},
  author={Xue, Ye and Wang, Yiheng and Shao, Xinhua and Yan, Qi and Zhang, Shutao and Chang, Tsung-Hui},
  journal={IEEE Transactions on Wireless Communications},
  volume={25},
  pages={17816--17830},
  year={2026},
  publisher={IEEE},
  doi={10.1109/TWC.2026.3696997},
  url={https://doi.org/10.1109/TWC.2026.3696997}
}

---

License

This project is licensed under the MIT License.