wimans

What it is

Public WiFi-CSI benchmark for multi-user sensing — the first CSI dataset built around 0–5 people present simultaneously (most CSI datasets assume a single actor). Acquired (2026-06-25) as the first real-data anchor for the thesis's occupancy→CSI-variance claim. Paper: huang2025_060d ↗ (ECCV 2024, arXiv 2402.09430).

Verified specs

• Hardware: Intel 5300, Linux 802.11n CSI tool (monitor mode), dual-band 2.4 / 5 GHz.
• Size: 11,286 samples × 3 s (~9.4 h); 3 real rooms (classroom, meeting room, empty room).
• CSI tensor / sample: 3000 × 3 × 3 × 30 (packets @1000 Hz × Tx-ant × Rx-ant × subcarriers). Raw wifi_csi/mat/*.mat; preprocessed amplitude wifi_csi/amp/*.npy (what the reduction uses).
• Labels (annotation.csv): environment, wifi_band, number_of_users 0–5, per-user location (A–E) + activity. Count split: 0 → 5.3 %, 1 → 31.6 %, 2–5 → 1782 each.

Acquisition & storage

• Source / license: Kaggle shuokanghuang/wimans — CC BY-NC-SA 4.0 (non-commercial research; do not redistribute in-repo).
• Canonical store: Hetzner S3 s3://monad-knowledge/datasets/wimans/ (the project object store; reductions read via the boto3 download-then-parse pattern). Local _attachments/datasets/wimans/ is a gitignored cache only — safe to delete once the S3 copy is verified.

What we proved with it (H4 → strong)

Reduction monad_knowledge/notebooks/python/csi_meneghello_zeroshot.py's sibling csi_wimans_zeroshot.py computes a hardware-agnostic pooled temporal CV (σ/μ over the packet axis) per sample and tests it against the real people-count:

• 5 GHz: ρ(count, CV) = +0.71, robust across all 3 rooms (+0.69/+0.79/+0.75); presence (empty-vs-occupied) AUC = 0.991; count-from-CV leave-one-room-out MAE ≈ 1.0 person (×1.1 cross-room). 2.4 GHz: presence-only (count ρ +0.28).
• This is the real-data evidence (with the meneghello independent-hardware confirmation) that moved occupancy-csi-variance (H4) to strong. Figures: _attachments/wimans-zeroshot/.

Notes

• The count is users present (not activity); we use the number_of_users field for crowd counting.
• Companion dataset on different hardware: meneghello-80mhz-csi (80 MHz Nexmon). Story: 2026-06-25 - Sim-to-real story — H4 strong, simulator bounded.