Style Transfer for Reconstruct 3D Human Pose from Single Image
Rong Tan, Jun Li, Zhiping Shi · 2021 3rd International Academic Exchange Conference on Science and Technology Innovation (IAECST) · 2021
In this paper, the image after the style transfer is used to train the neural network to learn texture information for 3D human shape reconstruction, and then the human pose which is extracted from Openpose is used as the ground truth for loss iteration. The auxiliary neural network can be obtained from the complex image Appropriate human parameters allow shape reconstruction through SMPL. the combination of iterative human parameters and training neural network to learn texture extraction information improves the accuracy of 3D human reconstruction. This method alleviates the low training speed and reduces the need for a large amount of supervision data to a certain extent. This work uses style transfer to enhance images and modify network structures that are validated on the dataset and robust.