widar

Description

The Widar family (Zheng et al., Tsinghua TNS) is one of the canonical WiFi CSI datasets for gesture and activity recognition with explicit cross-domain coverage. Widar3.0 (zheng2019_5389 ↗, MobiSys 2019) extends earlier versions with multi-orientation, multi-location, and multi-environment captures, and ships with a derived Body-coordinate Velocity Profile (BVP) representation that decouples gestures from spatial layout. It is the de facto reference for environment-independent CSI gesture recognition, and the canonical large-scale collection cited by the Tsinghua Hands-on Wireless Sensing with Wi-Fi tutorial (the WST site, Yang/Zhang/Chi/Zhang 2022, arXiv:2206.09532).

Modality / size

• Modality: WiFi CSI from intel-5300 NICs; six receivers per gesture for spatial diversity (the multi-receiver geometry is what makes BVP computable).
• Scale: ~258,000 gesture instances over ~8,620 minutes; ~75 spatial "domains" (the combinatorial product of rooms × positions × orientations).
• Subjects / scenarios: ~16–17 volunteers; a 6-gesture vocabulary used for BVP (Push&Pull, Sweep, Clap, Slide, Draw-Circle, Draw-Zigzag), with the full release spanning more gesture types; captured across multiple rooms, in-room positions, and body orientations for cross-domain splits.
• Derived representations / formats: raw CSI (.dat), DFS Doppler-frequency-spectrum (.mat, 6×121×T — 6 receivers × 121 frequency bins × time), and BVP (.mat, 20×20×T — velocity components in x and y over time in a body-centred frame).
• Labels: gesture class plus user / room / location / orientation metadata for cross-domain (zero-shot) evaluation.

Used by (papers)

• zheng2019_5389 ↗ — the originating paper; introduces Widar3.0 and the BVP feature.
• Cited extensively as a benchmark for cross-domain CSI gesture recognition (e.g. chen2024_b58b ↗).
• Used as a transfer-learning / domain-generalisation baseline in CSI sensing surveys (wang2026_2758 ↗, guarino2026_e72c ↗).
• Teaching anchor for WS501 Week 3 (Collecting CSI off commodity hardware) as the worked example of what exhaustive, multi-domain ground-truth collection costs and buys.

Notes

• Public release on the Widar3.0 project page; also distributed via IEEE DataPort, Tsinghua Disk, and Baidu Disk.
• The domain explosion is the key pedagogical point: each added factor (room, position, orientation, user) multiplies the hand-labelling burden, which is why so few datasets match its cross-domain coverage. The scripted, commanded-gesture protocol makes labels known by construction — the trade-off being that scripted gestures are not spontaneous behaviour.
• Contrast with the lab's BLE-assisted automated ground-truth direction (EXP-001 BLE-Assisted CSI Ground Truth Collection): automate the label to scale collection to occupancy, rather than hand-label a combinatorial grid.
• The earlier Widar (1.0/2.0) and Widar3.0 are folded together here because almost all current papers refer to Widar3.0; the alias list captures the older spellings.