A methodology for realistic human shape reconstruction from 2D images

This article presents a methodology for realistic human shape reconstruction from 2D images. The methodology involves adapting a state-of-the-art 3D human reshaping model and validating it on second the Anthropometric Survey of US Army Personnel (ANSUR II) data. A benchmark database of over 12,000 full-body silhouettes linked to a set of 21 body measurements was generated and used to train and test a deep-learning model to extract measurement information from a set of two images. A concatenated Convolutional Neural Network and Multilayer Perceptron model is trained to extract measurement information from two images while taking into account numerical inputs as constraints. Image augmentation is also used to improve accuracy and achieve generality. The results demonstrate that the proposed method is reliable and efficient, with most reconstructions achieving mean accuracy scores above 98%. Additionally, the body measurements extractor achieved a mean accuracy score of 97% for both male and female subjects. Overall, this methodology has significant potential for a range of applications, including the development of virtual try-on systems, personalized apparel design, and medical imaging technologies.