MAGIC-CSI: A Multi-Environment mmWave CSI Dataset for Micro-Gesture Recognition

We present MAGIC-CSI, a large-scale, high-resolution mmWave multiple-input multiple-output (MIMO) Channel State Information (CSI) dataset designed to advance fine grained sensing application like micro-gesture recognition under domain shifts. Unlike prior works that rely on mmWave-radar systems or sub- 6 GHz Wi-Fi CSI, MAGIC-CSI captures rich spatial-temporal features unique to mmWave MIMO systems, enabling fine-grained sensing in mmWave inte- grated sensing and communication (ISAC) systems. Moreover, the dataset of- fers a valuable opportunity to tackle a key challenge in wireless sensing\u2014domain adaptation\u2014by providing data collected from diverse subjects across multiple environments. The data is captured with m3MIMO\u2014an 8 \u00d7 8 fully digital mmWave MIMO testbed operating at 58 GHz with 1 GHz bandwidth\u2014 and includes 10 distinct micro-gestures performed by 2 subjects across 3 diverse in- door environments: a lab, a conference room, and a study room. Each gesture instance is captured through high-resolution CSI\u2014 also referred to as channel frequency response (CFR) with fine-grained temporal and spatial detail, and is precisely synchronized with video streams that serve as ground truth. Each subject performs all gesture classes 50 times across all environments, with each gesture lasting 3 seconds, resulting in a total of 3,000 labeled instances spanning multiple subjects and domains. All CSI samples are temporally aligned with labeled gesture annotations and corresponding video frames, enabling precise benchmarking including domain adaptive techniques. The dataset comprises both raw and preprocessed CSI data, where preprocessing involves a domain- adaptive pipeline featuring path loss compensation, frequency normalization, and entropy-guided filtering. The dataset is shared at:ieee dataport link.