BhavyaN/gsparc-datasets

GSpaRC Datasets

Datasets used in the paper "GSpaRC: Gaussian Splatting for Real-time Reconstruction of RF Channels".

• 📄 Paper: arXiv:2511.22793
• 🌐 Project website: https://nbhavyasai.github.io/GSpaRC/
• 💻 Code: https://github.com/Nbhavyasai/GSpaRC-WirelessGaussianSplatting

We evaluate GSpaRC on three RF datasets. Only the Sionna conference-room dataset is hosted in this repository (it was generated by us). The RFID and Argos datasets are publicly available from their original sources, linked below.

---

1. Sionna Conference-Room Dataset (hosted here)

Indoor spectrum dataset generated with the Sionna ray tracer. A single transmitter is fixed near the ceiling and 5,142 receiver positions are sampled across the room floor. Each receiver position yields one spatial spectrum image.

Files (under sionna_conference_room/):

File	Description
spectrum/00001.npy … 05142.npy	Per-position spatial spectrum (90 × 360 magnitude image, float32)
rx_pos.csv	Receiver positions, columns: x, y, z (5,142 rows)
gateway_info.yml	Transmitter position [0.0, 0.5, 3.0] and orientation quaternion

Splits used in the paper:

• 3,599 training positions / 1,543 test positions (random split)

Scene: indoor conference room, 14 m × 10 m × 4 m, with tables, chairs, glass partitions and irregular wall materials, producing rich multipath.

Loading example

import numpy as np
import pandas as pd
import yaml

# Receiver positions
rx_pos = pd.read_csv("sionna_conference_room/rx_pos.csv")  # columns: x,y,z

# Transmitter
with open("sionna_conference_room/gateway_info.yml") as f:
    info = yaml.safe_load(f)
tx_pos = info["gateway1"]["position"]  # [0.0, 0.5, 3.0]

# Spectrum at the i-th receiver position
spec = np.load(f"sionna_conference_room/spectrum/{i:05d}.npy")  # (90, 360)

---

2. RFID Spectrum Dataset (external)

Real-world dataset originally released with the NeRF² paper.

• Download: OneDrive link (provided by the NeRF² authors)
• Original repo: https://github.com/XPengZhao/NeRF2
• Setup: receiver fixed at 915 MHz with a 4×4 antenna array; an RFID-tag transmitter is placed at 6,123 distinct locations, each producing a spatial-spectrum image.
• Citation:

  Zhao, X., An, Z., Pan, Q., & Yang, L. (2023). NeRF²: Neural Radio-Frequency Radiance Fields. MobiCom 2023.

We use this dataset as-is — no preprocessing changes — and report results in the paper's RFID section.

---

3. Argos Massive-MIMO CSI Dataset (external)

Real-world outdoor CSI dataset from the Argos massive-MIMO platform.

• Download: OneDrive link (mirrored on the NeRF² authors' OneDrive)
• Original source: https://renew.rice.edu/dataset-argos.html
• Setup: 64-antenna base station at 2.4 GHz; 4,000 user positions; complex CSI on 26 subcarriers.
• Citation:

  Shepard, C., Yu, H., & Zhong, L. (2016). Understanding the Argos clean-slate massive MIMO platform. Rice University Technical Report.

For our experiments we follow the FIRE preprocessing (Liu et al., 2021) to obtain 3D receiver positions from the uplink measurements.

---

License

• The Sionna conference-room dataset (this repo) is released under CC-BY-4.0.
• The RFID and Argos datasets are governed by their respective original licenses — please refer to the source links above.

Citation

If you use any of these datasets together with GSpaRC, please cite:

@article{nukapotula2026gsparc,
  title   = {GSpaRC: Gaussian Splatting for Real-time Reconstruction of RF Channels},
  author  = {Nukapotula, Bhavya Sai and Tripathi, Rishabh and Pregler, Seth and Kalathil, Dileep and Shakkottai, Srinivas and Rappaport, Tedd},
  journal = {arXiv preprint arXiv:2511.22793},
  year    = {2026}
}