Domain-Adaptive Human Pose Estimation Without Source Data

EasyChair Preprint 14843

Abstract

Human pose estimation (HPE) has seen substantial progress through the application of supervised learning techniques using large labeled datasets. However, adapting these models to new environments often encounters challenges due to domain shift, particularly when source domain data is inaccessible. This article explores domain-adaptive human pose estimation without source data, focusing on innovative approaches to bridge the gap between source and target domains. We investigate several source-free domain adaptation techniques, including pseudo-labeling, entropy minimization, and adversarial learning, to enhance the model’s performance in unseen target domains. Through a detailed case study involving surveillance footage from an outdoor environment, we demonstrate the effectiveness of these techniques in improving pose estimation accuracy. Our findings reveal that source-free adaptation methods can significantly enhance model generalization, offering practical solutions for deploying HPE models in real-world applications where source data privacy or availability is a concern.

Keyphrases: Adversarial Learning, Domain Adaptation, Feature Alignment, Human Pose Estimation (HPE), Source-Free Adaptation, Surveillance Footage, entropy minimization, model generalization, pseudo-labeling, self-supervised learning

@booklet{EasyChair:14843,
  author    = {Alakitan Samad},
  title     = {Domain-Adaptive Human Pose Estimation Without Source Data},
  howpublished = {EasyChair Preprint 14843},
  year      = {EasyChair, 2024}}