Real2Sim2Real for Point Cloud-Based mmWave Link Quality Prediction: An Empirical Study

This paper demonstrates the feasibility of the real-to-sim-to-real (Real2Sim2Real) scheme for point cloud-based proactive millimeter-wave (mmWave) link quality prediction. To alleviate the mmWave line-of-sight blockage problem, proactive wireless network management based on predicted future link quality values leveraging computer vision-aided deep learning (DL) is an effective solution. However, a large dataset is required to train the high-accuracy DL-based prediction model. In this paper, we configure a Real2Sim2Real pipeline, which aims to automatically generate a realistic synthetic dataset, close to that obtained from real sensors, for pre-training by analyzing a small real-world dataset and computer simulation. Experimental results reveal that our method can reduce prediction errors even with a small sample of real-world datasets for fine-tuning compared with previous pre-training methods using manually created 3D simulation models.