TRI-POSE-Net: Adaptive 3D human pose estimation through selective kernel networks and self-supervision with trifocal...

Area of Science:

• Computer Vision
• 3D Human Pose Estimation

Background:

• Accurate 3D pose estimation for virtual entities is challenging due to the complexity of realistic movements.
• Conventional methods often fail to capture genuine avatar interactions in dynamic virtual environments.

Purpose of the Study:

• To introduce TRI-POSE-Net, a novel model for precise 3D pose estimation, particularly in scenarios with limited supervision.
• To enable 3D pose estimation from a single 2D RGB image.

Main Methods:

• The TRI-POSE-Net model integrates ResNet-50 with Selective Kernel Network (SKNet) blocks for efficient feature extraction.
• Trifocal tensors and trio-view geometry are employed to generate 3D ground truth poses from 2D poses, enhancing triangulation accuracy.

Main Results:

• The proposed approach achieved a Mean-Per-Joint-Position-Error (MPJPE) of 47.6 under self-supervision and 29.9 under full supervision on the HumanEva-I dataset.
• TRI-POSE-Net demonstrated strong performance, especially within the self-supervision paradigm, compared to existing methods.

Conclusions:

• TRI-POSE-Net provides an effective solution for 3D pose estimation in computer vision applications with limited supervision.
• The model's ability to generate accurate 3D poses from single 2D images advances virtual reality and avatar interaction technologies.