mm-fi

What it is

MM-Fi (Yang, Huang, Zhou, et al., MM-Fi: Multi-Modal Non-Intrusive 4D Human Dataset for Versatile Wireless Sensing, NeurIPS 2023 Datasets & Benchmarks) is a large multi-modal human dataset built for wireless human perception. It synchronises five sensing modalities against dense 3D human-pose ground truth, making it the reference cross-modal benchmark for WiFi/mmWave pose and action estimation. In the WiFlow repo it is the held-out cross-dataset used to test how a CSI pose model trained on one collection transfers to another.

Verified specs

• Subjects / scale: 40 subjects, 1080 sequences, >320,000 synchronised frames.
• Actions (27): 14 daily activities + 13 rehabilitation exercises (A01–A27).
• Environments (4): two room configurations (both ~8.5 m × 7.8 m), 10 participants each — the substrate for the dataset's cross-scene and cross-subject protocols.
• Modalities (5): RGB, depth, LiDAR point cloud, mmWave radar point cloud, and WiFi CSI.
• WiFi CSI: commodity 3×3 link geometry, 114 subcarriers, stored per-frame as .mat (CSIamp amplitude field); ~297 frames per sequence in the frame-level data unit (per the WiFlow loader cross_dataset_test/mmfi.py).
• Annotations: 2D/3D pose landmarks (17 keypoints), action label, 3D body position, estimated 3D dense pose.
• Official splits (protocol1/2/3 × random / cross-scene / cross-subject) shipped as a config.yaml; the WiFlow repo reuses this protocol machinery verbatim.

Acquisition & storage

• Access: freely available to researchers via the official release (loader + usage repo, project page); the raw bytes are behind a request/agreement form. Not yet mirrored to our S3 (no storage: block).
• Size caveat: full multi-modal release is large (hundreds of GB); the WiFi-CSI-only subset is the tractable pull and the only modality relevant here — mirror just wifi-csi/ per the datasets S3 contract, manifest keyed by subject / action / environment.

Why it matters here

• Highest-value acquisition of the three in the WiFlow repo. It is a large, well-curated, citable public benchmark with an explicit cross-scene / cross-subject design — exactly the domain-shift axis behind cross-geometry-feature-type-h1h6 and feature-type-transfer-campaign. It lets us test kinematic-vs-fingerprint feature transfer on real CSI instead of only synthetic Sionna traces.
• The multi-modal ground truth (LiDAR / mmWave alongside CSI) is a ready-made cross-modal label source, conceptually aligned with the BLE-assisted-ground-truth direction (ble-periodic-calibration) — here the "anchor" modality is radar/LiDAR rather than BLE.
• It sits in the same commodity-CSI family as wimans, widar, ntu-fi and xrf55 but is the only one of those pairing CSI with dense 3D pose across four environments — a genuine gap in the current vault dataset shelf.
• Limitation: single-occupant per capture, so like the pose sets it is not a multi-person count benchmark; its value here is cross-domain representation and label-transfer, not occupancy.