CUHK-S: A Privacy-Preserving Multimodal Dataset for Human Action Recognition

Dataset Description

CUHK-S is a privacy-preserving subset of the CUHK-X dataset, a large-scale multimodal benchmark for Human Action Recognition (HAR), Understanding (HAU), and Reasoning (HARn). CUHK-X was accepted at MobiSys 2026.

Compared to the full CUHK-X dataset, CUHK-S:

• Removes all RGB video to prevent facial identification
• Downscales all visual modalities to 320 × 240
• Selects 18 out of 30 participants while preserving full action coverage (40 categories)

Dataset Summary

Attribute	Value
Participants	18 (selected from 30 in CUHK-X)
Action Categories	40
Modalities	6 (Depth, IR, Thermal, IMU, Radar, Skeleton)
Visual Resolution	320 × 240
Total Size	~146 GB (18 zip files, one per participant)

Modalities

Modality	Format	Description
Depth	PNG (color)	Colorized depth maps from Vzense NYX 650
IR	PNG	Infrared images, robust to lighting changes
Thermal	PNG	Heat signature from thermal camera
IMU	CSV	5-sensor accelerometer/gyroscope/magnetometer
Radar	Binary	mmWave radar point cloud (TI Radar)
Skeleton	JSON/CSV	3D joint positions from pose estimation

Note: RGB video is intentionally excluded from CUHK-S to protect participant privacy.

Dataset Structure

Each participant's data is packaged as a zip file: CUHK-S_userN-userN.zip

CUHK-S/
├── HAR/                          # Human Action Recognition task
│   └── data/
│       ├── Depth_Color/          # Colorized depth frames (.png)
│       ├── IR/                   # Infrared frames (.png)
│       ├── Thermal/              # Thermal imaging frames (.png)
│       ├── Skeleton/             # Skeleton pose data
│       │   └── {action}/{user}/{session}/
│       │       ├── predictions/  # Keypoint JSON (.json) + overlay images (.jpg)
│       │       └── visualizations/
│       ├── IMU/                  # IMU sensor data (CSV)
│       │   └── {action}/{user}/{session}/
│       │       ├── up(LA+RA+C).csv   # Upper-body IMU (Left Arm + Right Arm + Chest)
│       │       └── down(LL+RL).csv   # Lower-body IMU (Left Leg + Right Leg)
│       └── Radar/                # mmWave radar data (CSV)
│           └── {action}/{user}/{session}/
│               └── radar_output_T{timestamp}.csv
│
├── HAU/                          # Human Action Understanding task
│   └── data/
│       ├── Depth/                # Visual modality clips as .mp4 video
│       ├── IR/
│       └── Thermal/
│           └── {user}/{session}/
│               └── {Modality}.mp4
│
├── HARn/                         # Human Action next-step Reasoning task
│   └── data/
│       ├── Depth/                # Video clips as .mp4
│       └── IR/
│           └── {action}/{user}/{session}/
│               └── Depth.mp4
│
└── source_data/                  # Raw source frames (with timestamps)
    └── data/
        ├── Depth_Color/          # Timestamped raw frames (.png)
        ├── IR/
        ├── Thermal/
        ├── Skeleton/
        ├── IMU/
        └── Radar/
            └── {user}/{session}/
                └── {Modality}_{timestamp}_{frameId}.png

Path naming convention:

Level	Meaning	Example
{action}	Action category with numeric prefix	10_Stir_drinks
{user}	Participant ID	user1
{session}	Scene–Environment–Trial index	2-1-1 (Scene 2, Indoor, Trial 1)

• HAR: Singular well-defined actions organized by action category, for traditional classification tasks
• HAU: Sequential action clips organized by user/session, for temporal and contextual understanding
• HARn: Sequential action clips organized by action/user/session, for next-action reasoning
• source_data: Original raw frames with full timestamps, before any task-level processing

IMU Sensor Layout

Five IMU sensors are placed on the body:

Sensor	Position	Channels (per sensor)
WTLA	Left Arm	Acc(X/Y/Z), Gyro(X/Y/Z), Mag(X/Y/Z)
WTC	Chest	Acc(X/Y/Z), Gyro(X/Y/Z), Mag(X/Y/Z)
WTRA	Right Arm	Acc(X/Y/Z), Gyro(X/Y/Z), Mag(X/Y/Z)
WTRL	Right Leg	Acc(X/Y/Z), Gyro(X/Y/Z), Mag(X/Y/Z)
WTLL	Left Leg	Acc(X/Y/Z), Gyro(X/Y/Z), Mag(X/Y/Z)

Benchmarks & Tasks

Task	Type	Metrics
Action Recognition	Classification	Accuracy, F1, Precision, Recall
Action Selection	Multiple Choice	Accuracy
Action Captioning	Generation	BLEU, METEOR
Emotion Analysis	Classification	Accuracy
Sequential Reordering	Ordering	Accuracy
Next Action Reasoning	Reasoning	Accuracy

Citation

If you use CUHK-S in your research, please cite:

@inproceedings{jiang2026cuhkx,
  title={CUHK-X: A Large-Scale Multimodal Dataset and Benchmark for Human Action Recognition, Understanding and Reasoning},
  author={Jiang, Siyang and others},
  booktitle={Proceedings of ACM MobiSys},
  year={2026}
}

Ethics & Privacy

We obtained approval from an Institutional Review Board (IRB) to conduct this study and collect data from human subjects.

Privacy measures in CUHK-S:

• No RGB video is included to prevent facial identification
• All visual modalities are downscaled to 320 × 240
• Participants are identified only by numeric IDs (e.g., user1, user2)
• No personally identifiable information is linked to individual records
• IMU, Radar, and Skeleton modalities do not capture visual appearance

License

Code is released under the MIT License. The dataset is available for non-commercial research under a Data Use Agreement (DUA) and is not redistributable. Our derived annotations/splits are released under CC BY 4.0.

Note: This dataset is designed for research and educational purposes. Please ensure compliance with your institution's ethics guidelines when using human activity data.

Contact

• Email: syjiang [AT] ie.cuhk.edu.hk
• Project Page: https://siyang-jiang.github.io/CUHK-X/
• Lab: CUHK AIoT Lab