Domain-Generalized 3D Human Pose Estimation via Dual Augmentation and Source-Free Adaptation

Abstract

The accurate estimation of three-dimensional human pose from monocular imagery remains a critical challenge in computer vision, primarily due to the severe domain shifts encountered when deploying models trained on controlled laboratory data to unconstrained, in-the-wild environments. Traditional domain adaptation techniques attempt to bridge this gap but typically require simultaneous access to both source and target datasets, raising significant data privacy and storage concerns. To address these limitations, this paper proposes a novel framework for Domain-Generalized 3D Human Pose Estimation utilizing a Dual Augmentation strategy coupled with a Source-Free Adaptation protocol. Our approach first enhances the generalization capability of the baseline model during the source training phase through a dual augmentation mechanism that systematically applies both spatial appearance transformations and kinematic structural perturbations. Subsequently, we introduce a source-free adaptation phase that aligns the pre-trained model to unseen target domains utilizing only unlabeled target data. This is achieved through the minimization of an unsupervised temporal consistency objective and a structural plausibility constraint, eliminating the dependency on the original source data. Comprehensive evaluations across multiple standard benchmarks demonstrate that our method significantly outperforms existing state-of-the-art techniques in cross-dataset generalization scenarios. The proposed framework not only mitigates the performance degradation typically associated with domain shift but also presents a highly practical solution for real-world deployment where source data access is restricted.