Han Sang-Hun ; Park Min-Gyu ; Yoon Ju Hong ; Kang Ju-Mi ; Park Young-Jae ; Jeon Hae-Gon

Abstract

High-quality 3D human body reconstruction requires high-fidelity andlarge-scale training data and appropriate network design that effectivelyexploits the high-resolution input images. To tackle these problems, we proposea simple yet effective 3D human digitization method called 2K2K, whichconstructs a large-scale 2K human dataset and infers 3D human models from 2Kresolution images. The proposed method separately recovers the global shape ofa human and its details. The low-resolution depth network predicts the globalstructure from a low-resolution image, and the part-wise image-to-normalnetwork predicts the details of the 3D human body structure. Thehigh-resolution depth network merges the global 3D shape and the detailedstructures to infer the high-resolution front and back side depth maps.Finally, an off-the-shelf mesh generator reconstructs the full 3D human model,which are available at https://github.com/SangHunHan92/2K2K. In addition, wealso provide 2,050 3D human models, including texture maps, 3D joints, and SMPLparameters for research purposes. In experiments, we demonstrate competitiveperformance over the recent works on various datasets.