Transfer Learning based Precise Pose Estimation with Insufficient Data

Abstract

With the recent advance in computer vision techniques and the growing utility of real-time human pose detection and tracking, deep learning-based pose estimation has been intensively studied in recent years. These studies rely on large-scale datasets of human pose images, for which expensive annotation jobs are required due to the complex spatial structure of pose keypoints. In this work, we present a transfer learning-based pose estimation model that leverages low-cost synthetic datasets and regressive domain adaptation, enabling the sample-efficient learning on precise human poses. In evaluation, we demonstrate that our model achieves the high accurate pose estimation on a dataset of golf swing images, which is targeted for a virtual golf coaching application.