Measurements-to-body: 3D human body reshaping based on anthropometric measurements

Abstract

Accurate 3D human models are useful for many applications in virtual fitting, ergonomics, film and television, and video games. However, due to the limitations of 3D scanners and body privacy, creating a virtual human body that accurately represents a specific human body is challenging. Therefore, reshaping 3D human bodies based on anthropometric measurements has received extensive attention. However, the existing methods have some drawbacks, such as the inability of the reshaped body to change its posture, the lack of a good link between the real and virtual measurements, and unreasonable anthropometry definitions. In this paper, we propose a new framework for reshaping the 3D human body using five easily available measurements: height, weight, chest, waist, and hip. First, the STAR model was used to fit the SPRING dataset to obtain the SPRING-fitted dataset, where the shape parameters of the STAR model are used to characterize each 3D human body. Second, optimizing the virtual measurement algorithm constructed a good link between real and virtual measurements. Then, the measurements of the human bodies in the SPRING-fitted dataset were extracted. Finally, the semantic reshaping of the 3D human body can be achieved by constructing a neural network model that uses the five measurements to predict 20 shape parameters. The results show that the human body reconstructed by our method can keep its size close to the real human body and conform to the shape of the real human body. Thus, it can meet the needs of the garment industry. In addition, the reshaped human body can be adjusted to different postures, which is beneficial to virtual fitting.