Top-Down System for Multi-Person 3D Absolute Pose Estimation from Monocular Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Recent advancements in 2D and 3D pose estimation exist.
• Real-world applications need 3D absolute poses and inter-person distances.
• Single-view 3D absolute pose estimation remains challenging and resource-intensive.

Purpose of the Study:

• To propose a real-time framework for multi-person 3D absolute pose estimation from a monocular camera.
• To overcome the limitations of existing resource-heavy systems.
• To enable accurate depth and distance estimations between individuals in a scene.

Main Methods:

• A novel top-down framework integrating human detection, 2D pose estimation, 3D root-relative pose reconstruction, and root depth estimation.
• Utilizes modified GAST-Net and RootNet architectures, forming the Root-GAST-Net system.
• Evaluated quantitatively and qualitatively on Human3.6M and MuPoTS-3D benchmark datasets.

Main Results:

• Root-GAST-Net achieves real-time performance at 15 fps on an Nvidia GeForce GTX 1080.
• Significantly outperforms current state-of-the-art methods on the MuPoTS-3D dataset across all metrics.
• Demonstrates efficiency and effectiveness through comprehensive evaluations.

Conclusions:

• The proposed Root-GAST-Net framework successfully addresses the challenge of multi-person 3D absolute pose estimation from monocular cameras.
• It offers a resource-efficient and high-performance solution for real-world applications requiring accurate 3D human pose information.
• The system provides a significant improvement over existing methods, paving the way for more advanced computer vision applications.