Knowledge Transfer from Human to Animal Pose Estimation Using Task Editing in Small-Scale Datasets

Abstract

With the rapid development of artificial intelligence, pose estimation has attracted attention not only for humans but also for animals. However, datasets for animals remain limited, and the high cost of annotation poses a major obstacle to practical deployment. This study investigates a method to construct animal pose estimation models with limited data by leveraging the abundance of pre-trained models for human pose estimation and applying task editing through Task Arithmetics. Specifically, difference vectors were derived from human and animal joint estimation tasks and applied as subtraction operations to examine the feasibility of knowledge transfer in low-sample domains. The experimental results suggest that smaller learning rates improve the effectiveness of Task Arithmetics, while differences in annotation policies have little impact on estimation accuracy.