Reconstructing 3D human avatars from monocular images

Abstract

Creating convincing representations of humans is a fundamental problem in both traditional arts and modern media. In our digital world, virtual avatars allow us to simulate and render the human body for a variety of applications, including movie production, sports, human-computer interaction, and medical sciences. However, capturing digital representations of a person’s shape, appearance, and motion is an expensive and time-consuming process which usually requires a lot of manual adjustments. With the advances in consumer-grade virtual reality devices, personalized virtual avatars became an essential part of interactive and immersive applications like telepresence and virtual try-on for online fashion shopping, thereby increasing the need for versatile easy-to-use self-digitization. In this chapter, we discuss a selection of recent acquisition methods for personalized human avatar reconstruction. In contrast to conventional setups, these fully-automatic approaches only use low-cost monocular video cameras to effectively fuse information from multiple points in time and realistically complete reconstructions from sparse observations. We address both straight-forward and sophisticated reconstruction methods focused on accuracy, simplicity, and usability to compare and provide insights into their visual fidelity and robustness.