Fine-tuning SMPL: A Framework for Highly Detailed Statistical Human Model Building

Nikolay Patakin, Ilia A. Petrov, Vladimir Guzov, Anton Konushin

Abstract: In this paper, we consider the task of statistical modeling of 3D human shape and pose. Current advances in computer graphics, 3D scanning and reconstruction technologies and emergence of new application areas for human body models raise the need not only in highly detailed human body model that can qualitatively represent details of human appearance, but also in a framework for fine-tuning this model based on newly collected data. We address both of these issues in our work by presenting a framework for fully automatic creation of highly detailed human body model based on SMPL model. The key features of our framework are mesh subdivision technique that increases the granularity of the model, modified Non-Rigid Deformation algorithm for smooth and precise registration of 3D scans and weighted registration process that allows controlling registration in low confidence areas of the 3D scan (holes, artifacts and voluminous haircuts). We propose and evaluate two body models with different detail level and show that even low detailed model outperform existing body models in terms of registration accuracy and cumulative relative variance.