YOLOR-comm-mmWave

YOLOR-comm-mmWave is a fine-tuned object detection model for BS identification for beam initialization to detect mmWave radio in one inference pass. The model is trained on imagery of Terragraph Sounders from Meta, deployed in indoor commercial spaces. Part of the YOLOR detector family used for the Look Once, Beam Twice mmWave V2X beam-management pipeline (SECON 2026).

Reference implementation for the paper:

Avhishek Biswas*, Apala Pramanik*, Eylem Ekici, Mehmet C. Vuran. "Look Once, Beam Twice: Camera-Primed Real-Time Double-Directional mmWave Beam Management for Vehicular Connectivity." (*equal contribution)

arXiv: https://doi.org/10.48550/arXiv.2605.05071

Quick links

• Paper (arXiv): https://doi.org/10.48550/arXiv.2605.05071
• Code: https://github.com/UNL-CPN-Lab/Look-Once-Beam-Twice
• Training pipeline: https://github.com/UNL-CPN-Lab/Look-Once-Beam-Twice/tree/main/YOLOR_Training

Architecture	YOLOv11x, 82-class output head (COCO 80 + 2 custom)
Initialization	stock yolo11x.pt
Schedule	200 epochs, cos_lr, close_mosaic=20, lr0=0.01
Training data	IndoorCommercialDataset, perceptual-hash deduped (cp_dedup.py, Hamming threshold = 1) — 1,631 train (kept from ~14,386 raw frames) / 1,798 val / 1,799 test
Custom classes	radio (id 80), mmWave radio (id 81)
Released checkpoint	last.pt

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(repo_id="cpnlab/YOLOR-comm-mmWave", filename="last.pt")
model = YOLO(weights)
results = model.predict("path/to/image.jpg", conf=0.25)

Class indices: 0–79 = COCO; 80 = radio; 81 = mmWave radio.

Training data

Code and Data: https://github.com/UNL-CPN-Lab/Look-Once-Beam-Twice

Citation

@inproceedings{biswas2026look,
  title     = {Look Once, Beam Twice: Camera-Primed Real-Time Double-Directional
               mmWave Beam Management for Vehicular Connectivity},
  author    = {Biswas, Avhishek and Pramanik, Apala and Ekici, Eylem and Vuran, Mehmet C.},
  booktitle = {Proc. IEEE SECON},
  year      = {2026}
}

Paper: https://doi.org/10.48550/arXiv.2605.05071

Contact

For questions about this model or the paper, contact the corresponding authors:

• Avhishek Biswas — abiswas3@huskers.unl.edu
• Apala Pramanik — apramanik2@huskers.unl.edu

Acknowledgments

Developed at the Cyber Physical Networking (CPN) Lab, School of Computing, University of Nebraska–Lincoln, in collaboration with The Ohio State University. Thanks to Sivers Semiconductors, Ettus Research, and the open-source Ultralytics, PyTorch, and Ettus UHD communities.