SDFPoseGraphNet: Spatial Deep Feature Pose Graph Network for 2D Hand Pose Estimation

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Hand pose estimation (HPE) is vital for human-computer interaction (HCI) and virtual reality (VR).
• Existing 2D HPE methods face challenges with hand dynamics and occlusions.
• Accurate feature extraction is critical for improving HPE performance.

Purpose of the Study:

• To introduce SDFPoseGraphNet, a novel framework for precise hand pose estimation.
• To enhance deep feature extraction from hand images using VGG-19 and spatial attention.
• To improve pose estimation accuracy by adaptively processing feature maps with a Pose Graph Model (PGM).

Main Methods:

• Utilized the VGG-19 architecture combined with spatial attention (SA) for refined feature map extraction.
• Integrated a Pose Graph Model (PGM) with First Inference Module (FIM) potentials and adaptive parameters for pose estimation.
• Developed an end-to-end trainable network optimizing all components for enhanced precision.

Main Results:

• SDFPoseGraphNet demonstrated superior performance compared to state-of-the-art methods.
• Achieved an average precision improvement of 7.49% over Convolution Pose Machine (CPM).
• Showed a 3.84% higher average precision compared to the Adaptive Graphical Model Network (AGMN).

Conclusions:

• SDFPoseGraphNet offers a significant advancement in hand pose estimation accuracy.
• The framework effectively addresses challenges posed by hand dynamics and occlusions.
• The proposed model provides a robust and precise solution for computer vision applications requiring accurate HPE.