A Two-Stage Augmentation Approach for Domain Generalization in 3D Human Pose Estimation

EasyChair Preprint 14841

Abstract

This article presents a novel two-stage augmentation approach designed to enhance domain generalization in 3D human pose estimation models. 3D human pose estimation, a critical task in computer vision, faces significant challenges when models are applied to unseen domains with different environmental conditions, lighting, and human appearances. Our proposed framework combines traditional data augmentation with feature augmentation techniques to create domain-invariant representations. In the first stage, geometric and color-based transformations are applied to the input data to improve robustness against domain shifts, while the second stage focuses on feature augmentation using adversarial learning and neural network-based transformations to ensure domain-invariant feature extraction. This two-stage approach leads to significant improvements in the model’s ability to generalize across diverse domains. Extensive experiments on popular 3D human pose datasets demonstrate the effectiveness of the proposed method, showing improved performance compared to traditional models. The study provides valuable insights into the potential of combining data and feature augmentations for domain generalization in computer vision tasks.

Keyphrases: 3D human pose estimation, Adversarial Learning, Convolutional Neural Networks (CNNs), Cross-domain robustness, Domain Generalization, Feature Augmentation, Human Computer Interaction, data augmentation, domain-invariant features, pose estimation

@booklet{EasyChair:14841,
  author    = {Kayode Sheriffdeen},
  title     = {A Two-Stage Augmentation Approach for Domain Generalization in 3D Human Pose Estimation},
  howpublished = {EasyChair Preprint 14841},
  year      = {EasyChair, 2024}}