Human Motion Video Generation for Pose Transfere

When generating human motion videos, pose migration faces two significant problems. The first is how to infer and transform the pose and pattern during the image migration process, and the second is how to ensure the smoothness and consistency of the generated video in the presence of noisy poses. To solve these problems, this paper proposes a deformable generation method that fuses pose features and reference image features. The method utilizes deformable convolution for implicit pose inference transformation to solve the difficulty of inference transformation in pose migration. Based on the deformable generation method, the motion vectors are further fused, and a deformable motion generation method that fuses the pose alignment features, motion vectors and reference image features is proposed. The module utilizes deformable convolution and parameter demodulation to fuse the above three so that the model generates continuous motion transformed video. The experimental results show that, compared with other mainstream models, the video generation model in this paper has a significant advantage in the video evaluation index FVD and outperforms most models in the image evaluation index.