Skip to content

WiMANS

  • Modality: WiFi Channel State Information (CSI)
  • Primary Tasks: Multi-user activity recognition, WiFi-based sensing, privacy-preserving HAR
  • Scale: Multi-user scenarios, multiple activity categories, varied spatial configurations
  • License: Research use
  • Access: Available through the official repository (see reference paper)

Summary

WiMANS is the first benchmark dataset for WiFi-based multi-user activity sensing presented at a top computer vision venue. By leveraging WiFi CSI signals instead of cameras, it enables human activity recognition without visual sensors, addressing privacy concerns in smart-home and workplace environments. The dataset captures simultaneous activities of multiple users, tackling a significantly harder problem than single-user WiFi sensing. This novel modality opens up research at the intersection of wireless communication and activity understanding.

Reference Paper

  • Authors et al. "WiMANS: A Benchmark Dataset for WiFi-based Multi-user Activity Sensing." ECCV, 2024. PDF

Benchmarks & Baselines

  • CNN/LSTM baselines on CSI - Classification accuracy reported for single-user and multi-user settings; WiMANS paper, ECCV 2024.
  • Cross-environment evaluation - Performance under different room layouts and device placements provided.
  • Official splits test generalization across spatial configurations and user combinations.

Tooling & Ecosystem

  • Official repository provides data loaders and baseline model implementations for CSI-based recognition.
  • CSI extraction tools compatible with commodity WiFi hardware (Intel 5300 NIC / Atheros-based).
  • Can be combined with vision-based datasets for multimodal sensor fusion research.

Known Challenges

  • WiFi CSI is sensitive to environmental changes (furniture movement, door state, new objects); models may not generalize across rooms without adaptation.
  • Multi-user signal separation is inherently ambiguous when users are spatially close.
  • Hardware-specific CSI formats vary across chipsets; preprocessing pipelines must account for antenna configuration and subcarrier count.
  • Relatively new modality for the CV community; fewer pretrained models and established best practices compared to video-based HAR.

Cite

@inproceedings{wimans2024,
  title     = {WiMANS: A Benchmark Dataset for WiFi-based Multi-user Activity Sensing},
  author    = {Huang, Shuokang and Li, Kaihan and You, Di and Chen, Yichong and Lin, Arvin and Liu, Siying and Li, Xiaohui and Cheung, Gene},
  booktitle = {Proceedings of the European Conference on Computer Vision},
  year      = {2024}
}