m monad-knowledge FIIT STU · open research vault
dataset

wipose

Also known as: WiPose, WiPose dataset, Person-in-WiFi 3D

What it is

WiPose (Jiang, Xue, Guo, et al., Towards 3D Human Pose Construction Using WiFi, ACM MobiCom 2020, DOI 10.1145/3372224.3380900) is the first dataset for 3D human-skeleton reconstruction from commodity WiFi CSI — joints on both limbs and torso, captured with a camera-derived 3D ground truth. In the WiFlow repo it is the second held-out cross-dataset (alongside mm-fi) used to probe cross-collection generalisation of CSI pose models.

Verified specs

Grounded in the WiFlow loader (cross_dataset_test/HPE-Li/wipose/wipose_dataset.py) + the paper:

  • Hardware: Intel 5300 NIC, 3 × 3 Tx/Rx antenna array → 9 links, 30 subcarriers per link — the classic csitool front-end (same NIC family as wimans and widar).
  • CSI tensor: .mat CSI field reshaped by the loader to (9 links × 30 subcarriers × 5 packets) per window (amplitude, per-channel mean/std normalisation constants baked into the repo).
  • Labels: 18 keypoints, each (x, y, confidence) — the loader keeps the 2D coordinates (mm → m scaling) plus a confidence channel; the source data carries the full 3D skeleton.
  • Split: pre-partitioned Train / Test directories of per-window .mat files (v7.3 / HDF5-backed, read with mat73).
  • Distribution artefact: released as wifipose_dataset.zip (unpacks to train_data / test_data); redistributed under several derivative repos (e.g. Person-in-WiFi-3D lineage).

Acquisition & storage

  • Access: research release, available on request / via mirrors rather than a single stable public portal — the canonical source is the MobiCom paper + authors; several GitHub re-hosts carry wifipose_dataset.zip. Not yet mirrored to our S3 (no storage: block).
  • If acquired: modest size; upload to datasets/wipose/ preserving the Train/Test tree, manifest keyed by split + activity, per the datasets contract. Note the v7.3 .mat format needs mat73/h5py, not scipy.io.loadmat.

Why it matters here

  • Lower priority than mm-fi but the only 3D CSI-pose set of the three and the smallest to acquire. Useful as a second real-CSI point for kinematic-feature transfer (feature-type-transfer-campaign, cross-geometry-feature-type-h1h6) and to sanity-check that claims hold across NIC front-ends (Intel-5300 3×3×30 here vs MM-Fi's 3×3×114).
  • Its 9-link × 30-subcarrier layout matches the widar/wimans Intel-5300 lineage, so tooling written for those transfers with little change.
  • Same task caveat: single-person pose, not occupancy/count — a representation-learning and cross-hardware reference, not a benchmark for the core counting hypotheses.