Enhancing Human Pose Estimation Across Domains Without Access to Source Data: Supplementary Information

EasyChair Preprint 14632

Abstract

This article explores advanced techniques for enhancing human pose estimation across various domains without the need for source data access. Traditional pose estimation models often rely on access to source data for fine-tuning and domain adaptation, which may not be feasible in many practical scenarios. To address this, we propose a novel approach that leverages supplementary information to improve domain generalization in human pose estimation. The approach involves using domain-invariant features and synthetic data generation to augment the training process. Experimental results demonstrate that the proposed method achieves significant improvements in pose estimation accuracy across different domains, outperforming baseline models that rely on direct source data access. This research contributes to advancing the field of human pose estimation by offering effective solutions for scenarios where access to source data is limited or unavailable.

Keyphrases: Benchmark Datasets, Convolutional Neural Networks (CNNs), Cross-Domain Performance, Generative Adversarial Networks (GANs), Mean Per Joint Position Error (MPJPE), data augmentation, domain-invariant features

@booklet{EasyChair:14632,
  author    = {Toluwani Bolu},
  title     = {Enhancing Human Pose Estimation Across Domains Without Access to Source Data: Supplementary Information},
  howpublished = {EasyChair Preprint 14632},
  year      = {EasyChair, 2024}}