Implementation of Cloth Estimation in 2D-3D Human Body Regression Model

Human body reconstruction is required to perform virtual try-on in cloth estimation. The reconstruction of 3D human body can provide the body shape and poses of the user. A fully expressive 3D human body regression model (ExPose) is used to reconstruct 3D human bodies from 2D images for the point-based modeling of the human clothing model. Subsequently, this model parameters is converted into SMPL-X and SMPL parameters from the given RGB image and inputs to the point-based modeling model along with the corresponding RGB image and clothing segmentation mask. A set of testing models in various poses is collected to evaluate the performance of point-based clothing estimation. A fitness measure is proposed to determine the fitness of the 3D body and outfit generated using SMPL-X and SMPL parameters. Intersection over Union and overlap percentage are also used to determine the set of parameter representing a 3D body in the point-based modeling model. This paper demonstrates that SMPL has better appearance retargeting performance and garment alignment capabilities for the point-based modeling model.