CSI2Depth: Spatio-Temporal Depth Images from Wi-Fi CSI Data via Transformer Networks and Conditional Generative Adversarial Networks

Abstract

Depth estimation is important for 3D reconstruction, robotics, and augmented reality. It is also increasingly relevant in Integrated Sensing and Communication (ISAC), where spatial awareness supports adaptive network optimization, beamforming, and digital twins. Traditional depth and RF-based sensors provide accurate measurements but often require active transmissions or specialized hardware. In contrast, Wi-Fi Channel State Information (CSI) enables passive sensing using existing wireless infrastructure. However, classical RF models struggle with multipath propagation, diffraction, and shadowing, limiting depth estimation in complex indoor environments. This study presents a novel proof-of-concept for generating depth maps from raw 5 GHz single-input multiple-output (SIMO) CSI data. A transformer encoder extracts spatio-temporal features from CSI amplitude and phase, which are processed by a conditional generative adversarial network (cGAN) to synthesize depth maps. Unlike prior CSI-based work focused on activity recognition or point cloud estimation, depth maps provide structured outputs that are computationally efficient and compatible with computer vision techniques. Evaluated on the public MM-Fi dataset, the model captures temporal depth variations associated with scene geometry and obstructions. The results demonstrate the feasibility of passive depth estimation from CSI under stable infrastructure with adequate signal quality and antenna diversity, contributing to the development of ISAC-enabled wireless systems. The code is available at: https://github.com/Arritmic/csi2depth.