A Survey of 3D Human Body Reconstruction from Single and Multiple Camera Views

Abstract

In today’s digital realm, where the metaverse, virtual reality, and video games converge, the importance of 3D human body reconstruction methods is undeniable. These techniques are useful for creating realistic avatars, lifelike animations, and immersive virtual environments. The problem of 3D model reconstruction remains subject to several challenges, including execution time of methods, the ability to reconstruct complex poses, and managing a variety of clothing. That is why human body reconstruction from one or multiple images or videos is an active area of research. The goal is to reconstruct the body surface geometry, and sometimes appearance, which refers to the textures. Reconstruction methods using parametric models require low computation time but may generate body shapes that are only roughly matching the input images. Thus, they may not be able to capture clothing and appearance details as accurately as non-parametric, photorealistic reconstruction methods, which are computationally intensive. The aim of this paper is to provide an overview of various reconstruction methods, ranging from those with minimal constraints (such as single-camera approaches) to those with more significant constraints (such as multi-camera techniques that require calibration data and prior segmentation mask extraction). To facilitate clarity, we will categorize the methods based on whether they produce a parametric or non-parametric model. This paper aims at (i) giving an overview of the existing methods, focusing on those published during the last two years, (ii) describing the requirements and constraints of each of these methods, (iii) presenting the metrics and datasets that are commonly used for the problem of human body reconstruction, (iv) and providing a classification and a comparison of these methods.