Single Shot Corrective CNN for Anatomically Correct 3D Hand Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Biomechanical Modeling

Background:

• Current 3D hand pose estimation methods prioritize accuracy over anatomical correctness.
• Existing datasets often contain ground truth poses with anatomical errors.

Purpose of the Study:

• To develop a novel 3D hand pose estimation model that enforces anatomical correctness at the architectural level.
• To address the limitations of post-facto filtering by integrating biomechanical constraints directly into the prediction model.

Main Methods:

• Introduction of the Single Shot Corrective CNN (SSC-CNN) architecture.
• Training and testing the SSC-CNN on HANDS2017 and MSRA datasets.
• Creation of Anatomically Error-Free (AEF) versions of benchmark datasets.

Main Results:

• SSC-CNN achieves accuracy comparable to state-of-the-art methods.
• The proposed model demonstrates zero anatomical errors, a significant improvement over previous approaches.
• AEF-HANDS2017 and AEF-MSRA datasets were created to mitigate existing ground truth errors.

Conclusions:

• SSC-CNN effectively integrates anatomical correctness into 3D hand pose estimation.
• The model offers a robust solution for accurate and biomechanically sound hand pose prediction.
• The development of AEF datasets is crucial for advancing research in anatomically correct hand pose estimation.