Indoor Over-The-Air samples of RF IQ samples in complex value

The following dataset includes two types of RF IQ samples: Synthetic and Indoor Over-The-Air datasets.This dataset has been used in a conference paper published in 2025 DySPAN: I Can’t Believe It’s Not Real: CV-MuSeNet: Complex-Valued Multi-Signal Segmentation.Paper abstract:The increasing congestion of the radio frequency spectrum presents challenges for efficient spectrum utilization. Cognitive radio systems enable dynamic spectrum access with the aid of recent innovations in neural networks. However, traditional real-valued neural networks (RVNNs) face difficulties in low signal-to-noise ratio (SNR) environments, as they were not specifically developed to capture essential wireless signal properties such as phase and amplitude. To address these limitations, this work presents CMuSeNet, a complex-valued multi-signal segmentation network for wideband spectrum sensing. Through extensive hyperparameter analysis, it is shown that a naive conversion of existing RVNNs into their complex-valued counterparts is not effective. Built on complex-valued neural networks (CVNNs) with a residual architecture, CMuSeNet introduces a complex-valued Fourier spectrum focal loss (CFL) and a complex plane intersection over union (CIoU) similarity metric to enhance training performance. Extensive evaluations on synthetic, indoor over-the-air, and real-world datasets show that CMuSeNet achieves an average accuracy of 98.98%-99.90%, improving by up to 9.2 percentage points (ppt) over its RVNN counterpart and consistently outperforming state of the art. Strikingly, CMuSeNet achieves the accuracy level of its RVNN counterpart in just two epochs, compared to the 27 epochs required for RVNN, while reducing training time by up to a 92.2% over the state of the art. The results highlight the effectiveness of complex-valued architectures in improving weak signal detection and training efficiency for spectrum sensing in challenging low-SNR environments.