Research map / Vision-language and multimodal / Human and embodied modeling
CSI-Bench: A Large-Scale In-the-Wild Dataset for Multi-task WiFi Sensing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CSI-Bench, a large in-the-wild WiFi CSI dataset from commercial edge devices across 26 homes/offices with 35 users, plus single-task and multi-task benchmarks and baselines.
Problem
WiFi sensing models generalize poorly because existing datasets are lab-collected, homogeneous in hardware, and session-based rather than continuous real-world recordings.
Approach
A large-scale, continuously recorded, in-the-wild CSI dataset spanning 16 commercial device types from five chipset vendors, with task-specific splits and a co-labeled multi-task subset (user ID, activity, proximity) evaluated with supervised and adapter-based multi-task baselines.
Key results
- Transformer multi-task training with LoRA adapters improves Human Activity Recognition accuracy from 75.40% to 88.06% and proximity recognition from 77.52% to 86.41% over task-specific training.
- Fall detection baselines reach about 94-95% accuracy on the medium split with ResNet-18 and LSTM at ~94.9% overall, while room-level localization reaches up to 100% for ResNet-18 and TimeSformer-1D.
- Under out-of-distribution splits, HAR accuracy drops to 54.7-61.8% and proximity recognition to 29.7-30.7%, showing a large generalization gap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-task joint training instead of task-specific training | Transformer · Human Activity Recognition | accuracy: 75.40 -> 88.06; +12.66 | Helped |
| Use multi-task joint training instead of task-specific training | Transformer · Human Activity Recognition | F1: 75.49 -> 86.00; +10.51 | Helped |
| Use multi-task joint training instead of task-specific training | Transformer · User Identification | accuracy: 99.51 -> 99.55; +0.04 | No clear effect |
| Use multi-task joint training instead of task-specific training | Transformer · User Identification | F1: 99.51 -> 99.70; +0.19 | No clear effect |
| Use multi-task joint training instead of task-specific training | Transformer · Proximity Recognition | accuracy: 77.52 -> 86.41; +8.89 | Helped |
| Use multi-task joint training instead of task-specific training | Transformer · Proximity Recognition | F1: 77.35 -> 87.09; +9.74 | Helped |
| Evaluate on unseen domains rather than in-distribution data | Transformer multi-task model · Human Activity Recognition, cross-device | accuracy | Hurt |
| Evaluate on unseen domains rather than in-distribution data | Transformer multi-task model · Human Activity Recognition, cross-environment and cross-user | accuracy | Hurt |
Surprising findings
- Multi-task joint training with shared backbone outperformed task-specific training on HAR and proximity, despite reduced parameters.
- Proximity recognition under cross-device, cross-environment, and cross-user shifts falls to near 30% accuracy.
Limitations
- Amplitude-only CSI features; phase unused due to cross-platform instability.
- Benchmark limited to classification tasks; regression and long-term tracking not included.
- Large OOD performance drops indicate weak cross-device, cross-environment, and cross-user generalization.
Experimental setup
Datasets: CSI-Bench
Benchmarks: CSI-Bench
Metrics: accuracy weighted F1
Baselines: MLP LSTM ResNet-18 Vision Transformer PatchTST TimeSformer-1D
Builds on: LoRA PatchTST Vision Transformer
Closest papers by meaning
- WiFi CSI Based Temporal Activity Detection via Dual Pyramid Network (AAAI 2025 · similarity 0.88)
- ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking (AAAI 2025 · similarity 0.87)
- Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.84)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.81)
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.81)
- Multitarget Device-Free Localization via Cross-Domain Wi-Fi RSS Training Data and Attentional Prior Fusion (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.