3D Human Pose and Shape Estimation Based on SMPL Model
Yuanyuan Song, Hui Zhou · 2024
D human pose and body shape estimation is a hot research topic in computer vision, and there are many optimization-based methods that rely only on the input parameters to obtain the optimal solutions for pose and body shape through multiple iterations. However, such methods often ignore the spatial information that may be embedded in the images, and we propose an effective method for reconstructing 3D human body models from a single RGB image. Our method effectively combines a priori information with picture spatial information. The feature information of the image is first obtained by a convolutional neural network. Then, the image features are combined with the joint a priori information to guide the generation of optimal solutions for the parameters, and we propose an improved method based on the optimization method, which enables the parameter updating process to effectively combine the image spatial features, and to obtain the estimation results matching the objectives. The whole process combines the advantages of the CNN method with the optimization-based method to guide the parameter update through the spatial feature information extracted from the CNN. The results of quantitative and qualitative experiments conducted on public datasets mark that our method outperforms previous methods in many aspects.