A dataset for human pose estimation using infrared imaging and low-cost reflective markers

Overview IRPose16 is a high-resolution infrared imaging dataset designed to support reproducible research in human pose estimation, motion tracking, and rehabilitation under constrained or low-visibility environments. The dataset combines low-cost infrared sensing with a reflective marker system to enable robust and consistent keypoint detection independent of ambient lighting conditions. The dataset contains 101,891 frames collected from 10 subjects, each instrumented with 16 custom-designed reflective markers placed at anatomically consistent keypoints. Every frame provides 100% complete keypoint coverage, resulting in approximately 1.63 million validated annotations, making IRPose16 a reliable resource for training, evaluation, and benchmarking of pose estimation models. This dataset is distributed with open-access annotations and controlled-access raw video data, enabling both usability and responsible data sharing. Data Collection Setup Recordings were acquired using an Orbbec Femto Bolt IR/RGB camera system, configured at a fixed distance and calibrated using a checkerboard-based geometric correction procedure. Both infrared (IR) and RGB streams were captured simultaneously: IR modality was used for precise marker localization RGB modality was retained for visualization and downstream applications Data acquisition was performed in a controlled indoor environment with uniform illumination and a structured green-screen background to minimize noise and ensure reproducibility across sessions. Marker System and Imaging Methodology IRPose16 employs a custom-designed reflective marker system optimized for infrared imaging within the 700–900 nm spectrum. Markers were fabricated using retroreflective materials to ensure high contrast under IR illumination. Markers were placed according to biomechanically consistent anatomical landmarks, aligned with the MPII 16-keypoint format, ensuring compatibility with standard human pose estimation benchmarks. The modular marker design enables: Consistent placement across subjects Robust detection under motion Adaptability to varying body types and clothing conditions Annotation Pipeline and Techniques The dataset annotations were generated using a structured multi-stage pipeline designed to ensure high spatial precision and temporal consistency: Blob Detection:Initial marker localization using OpenCV’s SimpleBlobDetector Temporal Linking:Nearest-neighbor tracking to maintain keypoint identity across frames Kalman Filtering:Temporal smoothing and interpolation to correct missing or noisy detections Multi-stage Validation: Deep autoencoder-based reconstruction PCA-based pose constraint enforcement Gaussian Mixture Model (GMM) likelihood screening IR–RGB Alignment:Homography-based projection using ArUco markers to map IR keypoints into the RGB coordinate space This pipeline ensures robust annotations, even under fast motion and partial occlusion scenarios. Content and Motion Diversity Each participant performed a set of controlled full-body motion sequences, including: Arm raises and rotations Bending and stretching movements Squats and lower-body exercises Leg rotations and balance movements Stair-climbing and cycling simulations These actions introduce significant pose variability, making IRPose16 suitable for: Human pose estimation Action recognition Temporal modeling Biomechanical and rehabilitation analysis Reproducibility All aspects of the dataset—including hardware configuration, calibration procedures, marker design, motion protocols, and annotation pipelines—are fully documented. The system is designed to be: Low-cost Sensor-agnostic Easily replicable This enables researchers to reproduce, extend, or adapt the dataset in different environments with minimal setup complexity. Data Access and Usage Conditions Processed annotations and non-identifiable data are openly accessible.Raw video data containing identifiable participants are provided under controlled access and require users to: Agree to a Data Use Agreement (DUA) Obtain approval from the authors prior to access This ensures responsible data usage while maintaining participant privacy. Ethics and Approval All procedures involving human participants were conducted in accordance with institutional ethical guidelines. Participants were recruited as voluntary adult subjects and provided written informed consent for data collection and controlled data sharing. To protect participant privacy, access to identifiable data is restricted and governed by a Data Use Agreement (DUA). Versioning and Availability This dataset is hosted on Zenodo and is associated with a persistent DOI, ensuring long-term accessibility, citation, and version control. Updates and future extensions of the dataset will be managed through versioned releases.