A Baseline for Cross-Database 3D Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Current 3D human pose estimation models struggle with real-world variability due to limited dataset diversity.
• Existing datasets (HumanEva-I, Human3.6M, Panoptic Studio) exhibit biases in subjects, poses, cameras, and lighting.
• Developing robust systems for arbitrary conditions ('in-the-wild') is crucial for practical applications.

Purpose of the Study:

• To investigate dataset biases in vision-based 3D human pose estimation.
• To improve the generalization capabilities of 3D human pose estimation models across diverse datasets and conditions.
• To advance towards 'in-the-wild' 3D human pose estimation.

Main Methods:

• Harmonized differing skeleton joint definitions across HumanEva-I, Human3.6M, and Panoptic Studio datasets.
• Proposed and applied a scale normalization method to improve cross-dataset generalization.
• Conducted cross-database experiments using a surrogate for estimating in-the-wild performance.
• Investigated effects of camera count, multi-dataset training, anatomy-based pose validation, and OpenPose integration.

Main Results:

• Joint harmonization significantly reduced biases and systematic errors in cross-database experiments.
• Scale normalization substantially improved generalization across camera viewpoints, subjects, and datasets.
• Augmenting virtual cameras enhanced both cross-database and in-database generalization.
• Identified inherent dataset biases that remain uncompensated, highlighting the need for more diverse datasets.

Conclusions:

• Joint harmonization and scale normalization are effective techniques for improving 3D human pose estimation generalization.
• While improvements were made, significant dataset biases persist, necessitating the creation of more diverse datasets.
• Future work should focus on developing methods to overcome remaining biases and create truly 'in-the-wild' capable systems.