Improvement of Pose Transfer by Introducing 3D Pose Estimation
Mahiro Watanabe, Ikuko Shimizu · 2023
Pose Transfer is a technique for generating images in which the posture of a person in an image is converted to a specified pose without changing the person’s appearance. PISE, a conventional Pose Transfer technique, has an advantage to generate high quality images based on clothes parsing analysis and 2D pose input, but the generated images may have unnatural postures. One of the reasons for this is that the front-back relationship of the bone joints is unclear because the input pose is in 2D. In this paper, we propose a PISE based Pose Transfer method to improve the accuracy of generated images by introducing 3D poses as input for PISE. MeTRAbs, a method for estimating 3D poses from real images, is used in our method. The encoder structure of PISE is also modified because the dimension of the input feature is increased. Experimental results show that the results obtained with our method are better than those obtained with conventional methods.