Enhancing Human Motion Transfer with Gromov-Wasserstein and Perceptual Loss Functions

First, it is important to note that the existing methods of human motion transfer have not utilized the Gromov-Wasserstein loss function. In our IJCAI 2022 paper \cite{liu2022copy}, we propose a motion transfer method that incorporates the Gromov-Wasserstein loss function. However, as you have mentioned, the current practice in human motion transfer involves the use of perceptual loss function. In this journal paper, we have discovered that combining the Gromov-Wasserstein loss function with the perceptual loss function yields superior results in human motion transfer. The Gromov-Wasserstein loss function is capable of preserving the distance-structure of the feature space, which is distinct from the conventional pixel-wise L2 loss. To further enhance the image reconstruction details, we introduce the perceptual loss function, which complements the Gromov-Wasserstein loss function, into the network. Consequently, the joint utilization of the Gromov-Wasserstein loss and perceptual loss facilitates the generation of realistic appearances.