SARN: Shifted Attention Regression Network for 3D Hand Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Biomedical Engineering

Background:

• Hand pose estimation (HPE) is crucial for hand function assessment and rehabilitation.
• Current HPE methods face challenges in accuracy and convenience during functional tasks.
• Accurate 3D hand pose data is essential for understanding hand biomechanics and developing assistive technologies.

Purpose of the Study:

• To propose a novel shifted attention regression network (SARN) for accurate 3D hand pose estimation (HPE) from depth images.
• To evaluate SARN's performance on established 3D hand pose datasets.
• To assess SARN's feasibility for characterizing functional hand movements, such as in neurological assessments.

Main Methods:

• Developed a novel shifted attention regression network (SARN) architecture for HPE.
• SARN predicts spatial relationships between depth image points and hand keypoints to infer 3D positions.
• Validated the method on NYU, ICVL, and MSRA 3D hand pose datasets and a custom finger tapping dataset.

Main Results:

• Achieved state-of-the-art performance on NYU (7.32 mm), ICVL (5.91 mm), and MSRA (7.17 mm) datasets.
• Demonstrated a mean error of 2.99 mm on a custom finger tapping dataset.
• Showcased comparable performance in distance, velocity, and acceleration metrics for finger movement analysis.

Conclusions:

• The proposed SARN method significantly advances 3D hand pose estimation accuracy and efficiency.
• SARN shows strong potential for real-world applications in functional hand assessment and rehabilitation.
• The method's effectiveness in characterizing hand movements opens avenues for clinical diagnostics and human-computer interaction.