Auto-Refining Reconstruction Algorithm for Recreation of Limited Angle Humanoid Depth Data.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Current 3D object reconstruction methods primarily focus on single, static synthetic objects.
• There is a significant need for techniques that can reconstruct morphing objects in dynamic scenes without external influence.
• Creating ground truths for real-world dynamic object reconstruction is a time-consuming process.

Purpose of the Study:

• To propose a novel deep adversarial neural network architecture for full human body posture reconstruction.
• To address the challenge of reconstructing dynamic human poses from real-world depth sensor data.
• To overcome the limitations of existing methods by enabling reconstruction without external influence and reducing reliance on time-consuming ground truth creation.

Main Methods:

• A three-staged deep adversarial neural network was developed.
• The network is designed to denoise and refine real-world depth sensor input.
• The approach focuses on full human body posture reconstruction.

Main Results:

• Achieved Earth Mover distance of 0.059 and Chamfer distance of 0.079 on synthetic datasets, demonstrating competitive performance.
• Successfully reconstructed human body posture from maskless, real-world depth frames.
• Visual inspection confirmed the network's ability to handle common depth sensor noise, excluding significant depth field deformities.

Conclusions:

• The proposed deep adversarial network offers a robust solution for 3D human body posture reconstruction from noisy depth data.
• The method demonstrates effectiveness in dynamic scenes and reduces the need for extensive real-world ground truth data.
• This approach advances the field of 3D reconstruction for dynamic human subjects.