SUMediPose: A 2D-3D pose estimation dataset.

Area of Science:

• Biomechanics
• Computer Vision
• Machine Learning

Background:

• Biomechanical movement analysis is vital but technologically expensive and inaccessible.
• Markerless motion capture using Pose Estimation (PE) offers alternatives but faces accuracy and data challenges.
• Existing 3D pose estimation methods struggle with hardware complexity and lack of reliable datasets.

Purpose of the Study:

• To introduce a comprehensive multimodal dataset for advancing 2D and 3D human pose estimation.
• To overcome the limitations of current markerless motion capture techniques.
• To provide a robust dataset for training and validating pose estimation models in biomechanics.

Main Methods:

• Developed a multimodal dataset with 3,444 recordings and over 3.8 million keypoint coordinates.
• Utilized a custom multi-RGB-camera system (6 cameras) for 360° capture and Vicon for ground truth 3D motion data.
• Synchronized marker-based motion capture with multi-camera RGB data, projecting 3D ground truth into 2D image space.

Main Results:

• The dataset contains 2,896,943 image frames from 28 participants performing 8 actions at 3 speeds.
• Includes precise 3D and 2D marker-based keypoint data, along with camera parameters for accurate projections.
• Offers a unique resource for both 2D and 3D pose estimation tasks, validated by professional marker placement.

Conclusions:

• The introduced dataset significantly enhances the potential for accurate and accessible biomechanical movement analysis.
• It addresses the critical need for large-scale, anatomically accurate data in 3D pose estimation research.
• This resource facilitates the development of more reliable markerless motion capture systems for medical and sports applications.