pku-pcni-lab/Multi-modal_dataset_named_SynthSoM-Twin

SynthSoM-Twin: Digital Twin SoM Dataset

📌 Overview

Based on the construction architecture of the SynthSoM dataset, the SynthSoM-Twin dataset further completes the digital twin reconstruction of the real-world scenarios in DeepSense 6G, achieving spatio-temporal consistency between the physical world and the simulation environment. The dataset covers 2 typical V2I scenarios, including urban and suburban scenarios. Additionally, data acquisition and construction comprehensively consider 3 types of scenario conditions: multi-frequency bands, multi-perspectives, and multi-vehicle traffic densities. The reconstruction content of the digital twin scenarios in the dataset includes two parts: static scenes and dynamic objects. For static scenes, real 3D reconstruction models are used to restore static scenes; for dynamic objects, multi-modal detection and tracking algorithms are used to restore dynamic object types and trajectories. The constructed SynthSoM-Twin dataset contains spatio-temporally synchronized 70,000 RGB images, 70,000 depth maps, 70,000 LiDAR point clouds, 20,000 millimeter-wave radar point clouds, and 70,000 large- and small-scale channel fading samples. Experimental results show that based on the constructed SynthSoM-Twin dataset, injecting less than 15% of real-world data can achieve a balance between model performance and real-world data cost, indicating that a suitable digital twin dataset can reduce measurement costs by an average of 85%.

---

📜 Related paper

J. Chen, Z. Huang, X. Cai, X. Cheng, and L. Yang, “SynthSoM-Twin: A multi-modal sensing-communication digital-twin dataset for Sim2Real transfer via Synesthesia of Machines,” IEEE Transactions on Machine Learning in Communications and Networking, submitted for publication, 2025.