Domain Adaptation for Cross-Target Pose Estimation via Task Editing

Abstract

Achieving high-precision pose estimation remains dependent on the labor-intensive annotation of keypoints, which poses a significant challenge in terms of human resource cost. This study aims to mitigate such costs by investigating the applicability of knowledge acquired in a source domain to a previously unseen target domain. Leveraging model editing techniques based on Task Vectors, we transfer the knowledge obtained from human pose estimation to the task of animal keypoint estimation. The edited models are evaluated in terms of both prediction accuracy and tendency, thereby assessing the effectiveness of domain adaptation across different target categories.