Research On Pose Migration Method Based On Adversary Network

Existing motion transfer algorithms learn a mapping from the skeleton or 3D models to real images,which means that the results generated by these algorithms are affected by the overall pose.Under the conditions of the short internet video,these methods tend to overfitting and losing a lot of detail information.For a pose,it can be broken up into several independent subparts.In this work,we use a new training method that allows the network to generate images from local to global.Our model can transfer motions between two videos with a small amount of data while being robust to a new posture.For each submodule,we add spatial and texture control constraints so that they still retain the constraints of the whole individual identity when generating the local region.Our main contributions are as follows: 1.We divide the motion transfer task into three sub-modules,each sub-module corresponding to a different task.2.We give the optimized loss function to the action migration algorithm to improve the stability of training.3.We optimized the network structure of the action migration task to verify the effect of this structure in multi-data fusion.Finally,we give the results of multiple network short video action migrations,and our method can maintain high robustness with a small amount of data.