A Bimodal Augmentation Model for Domain Adaptation in 3D Human Pose Estimation

EasyChair Preprint 14845

Abstract

3D human pose estimation (3D HPE) is a vital technology with applications spanning motion capture, augmented reality, and human-computer interaction. Despite its advancements, domain adaptation remains a challenge due to the variability in data sources and environments. This article introduces a bimodal augmentation model designed to address these challenges by enhancing domain adaptation in 3D HPE. Our model integrates two augmentation strategies—image-based and feature-based—aiming to bridge the gap between training and testing domains. Through a combination of synthetic data generation and feature transformation techniques, the proposed approach enhances the generalization capabilities of pose estimation algorithms across diverse domains. Experimental evaluations demonstrate that the bimodal augmentation model surpasses traditional single-modal methods in accuracy and robustness across various testing scenarios. This research offers a novel framework for improving the adaptability of 3D human pose estimation systems and provides insights into future directions for advancing domain generalization.

Keyphrases: 3D human pose estimation, Bimodal Augmentation, Cross-Domain Performance, Domain Adaptation, Domain Generalization, Feature-Based Augmentation, Image-Based Augmentation, Pose Estimation Algorithms, Robustness in Computer Vision, Synthetic Data Generation

@booklet{EasyChair:14845,
  author    = {John Owen},
  title     = {A Bimodal Augmentation Model for Domain Adaptation in 3D Human Pose Estimation},
  howpublished = {EasyChair Preprint 14845},
  year      = {EasyChair, 2024}}