HUMANNET-A Two-Tiered Deep Neural Network Architecture for Self-Occluding Humanoid Pose Reconstruction.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Current 3D reconstruction methods primarily focus on static objects.
• Existing approaches fail in dynamic and morphing scene reconstruction, limiting real-world applications.

Purpose of the Study:

• To develop a novel deep neural network architecture for reconstructing self-obstructed, human-like morphing shapes from depth frames.
• To address the limitations of static reconstruction in dynamic environments.

Main Methods:

• A two-tiered deep neural network was designed.
• The network utilizes depth frames and camera intrinsic parameters for reconstruction.
• Testing was conducted on a custom dataset derived from AMASS and MoVi datasets.

Main Results:

• The first tier achieved a Jaccard's Index of 0.7907 for region of interest extraction.
• The second tier demonstrated strong performance with an Earth Mover's Distance of 0.0256 and Chamfer Distance of 0.276.
• Subjective analysis confirmed the network's capability to reconstruct limb positions from minimal object details.

Conclusions:

• The proposed two-tiered network effectively reconstructs dynamic, morphing human shapes.
• The method shows significant promise for real-world applications requiring dynamic scene understanding.
• The network exhibits strong predictive capabilities, even with sparse input data.