This repository accompanies the paper:
Robust In-Bed Human Pose and Shape Estimation from Pressure Images with Clinical Awareness
We introduce CHIPS, a clinically grounded dataset for pressure-based in-bed human pose and shape estimation, and BA-PIHMR, a robust reconstruction framework designed for clinically realistic bed configurations.
- Dataset access is controlled and subject to approval, approved de-identified subsets may be provided upon request.
- The release scope of the dataset will follow ethical approval, privacy protection requirements, and institutional data governance policies.
To request access, please complete and sign the following form:
CHIPS Dataset Access Application and Data Use Agreement
Submit the completed PDF to:
Chenfang Fang
Email: fang_chenfang@mail.ustc.edu.cn
Suggested email subject:
CHIPS Dataset Application - [PI Full Name] - [Institution]
Estimating the 3D pose and shape of in-bed patients is important for clinical applications such as pressure ulcer prevention, sleep quality monitoring, and patient behavior analysis.
Pressure sensing provides a privacy-preserving and unobtrusive alternative to RGB cameras and wearable sensors. However, existing pressure-based datasets are often collected in simplified environments and do not fully reflect realistic clinical settings.
CHIPS addresses this limitation by collecting pressure data in an ICU-simulated environment with clinically relevant bed configurations, including head-of-bed elevation, multiple pillows, bedsheets, and medical-grade airbed effects.
CHIPS is designed for pressure-based 3D human pose and shape estimation in clinically realistic in-bed scenarios.
The dataset contains:
- 207,508 synchronized frames
- 16 subjects
- 50 clinically relevant ICU-style scenarios
- De-identified pressure images
- Derived 3D pose and SMPL-based shape annotations, where applicable
- Clinical risk-aware evaluation metadata for RW-MPJPE
The data collection setup used synchronized pressure sensing and multi-view RGB capture for annotation generation. The standard requestable/releasable dataset includes de-identified pressure images and derived annotations. Raw multi-view RGB recordings are not included in the standard release unless separately approved.
We propose BA-PIHMR, a Blur-Aware Pressure-Image-supported Human Mesh Reconstruction framework.
The model includes:
- A pressure image encoder for spatial feature extraction
- Bed-State FiLM for adapting to different bed configurations
- Blur-Aware Attention for handling pressure blur and ambiguous contact regions
- A temporal Transformer for sequence modeling
- An SMPL regressor for 3D human pose and shape recovery
We introduce RW-MPJPE: Risk-Weighted Mean Per Joint Position Error.
RW-MPJPE assigns higher evaluation importance to joints and anatomical regions that are more clinically relevant to pressure ulcer risk. The weighting schema was defined with input from ICU nursing experts.
The RW-MPJPE evaluation protocol and related metadata are provided to approved dataset users as part of the controlled-access dataset package.
Use of the CHIPS dataset is restricted to approved non-commercial academic research purposes only.
Researchers must not:
- Attempt to re-identify, de-anonymize, or contact any individual represented in the dataset.
- Redistribute, sublicense, sell, lease, or transfer the dataset to any third party.
- Use the dataset for commercial products, commercial services, or revenue-generating activities.
- Use the dataset for clinical diagnosis, clinical decision-making, patient monitoring, patient care, or any regulated medical use.
- Store the dataset on publicly accessible servers or unapproved cloud services.
- Use the dataset to train generative AI systems, large language models, or other machine learning systems for commercial deployment without a separate written agreement.
All dataset access requests must be reviewed and approved by the data provider. Approved users must follow the terms in the CHIPS Dataset Access Application and Data Use Agreement.
The repository code and documentation are released under the MIT License, unless otherwise stated.
The CHIPS dataset is not released under the MIT License. Dataset access and use are governed separately by the CHIPS Dataset Access Application and Data Use Agreement.
This study was performed in line with the principles of the Declaration of Helsinki.
Ethical approval was granted by the Medical Research Ethics Committee of The First Affiliated Hospital of the University of Science and Technology of China (USTC).
We thank the nurses from the Department of Intensive Care Unit, The First Affiliated Hospital of USTC, for their clinical expertise and support in defining the RW-MPJPE evaluation protocol and clinically relevant data collection scenarios.
This work is supported by the Research Funds of the Centre for Leading Medicine and Advanced Technologies of IHM.
The formal citation will be added after publication.
If you use the CHIPS dataset, BA-PIHMR method, or RW-MPJPE evaluation protocol, please cite the paper once available and acknowledge the dataset as follows:
This research used the CHIPS dataset (Robust In-Bed Human Pose and Shape Estimation from Pressure Images with Clinical Awareness), provided by Chenfang Fang, University of Science and Technology of China (USTC).
For dataset access, questions, or reporting potential privacy/security concerns, please contact:
Chenfang Fang Email: fang_chenfang@mail.ustc.edu.cn