Enhanced Human Pose Retargeting Through Video Frame Deblurring

초록

Human pose retargeting is a technology that gener-ates a new video by inputting both a skeleton video and an image containing a random person. Human pose retargeting involves estimating a 2D pose from each video frame to create a skeleton map. In videos with intense movements, such as dancing or exercising, many frames are inevitably blurred. These blurred frames hinder accurate 2D pose estimation. To address this issue, video motion deblurring technology is employed, using the BiT block to share features between frames and output sharpened frames. By performing deblurring on all frames of the video intended for human pose retargeting, 2D pose estimation can be carried out more accurately. In this study, we improved the accuracy of MagicPose's Pose ControlNet by enabling more precise pose estimation, and we succeeded in generating more natural videos. © 2025 IEEE.