Review of single-image 3D human reconstruction based on deep learning

Research progress and development tendencies of deep-learning-based single-image 3 dimensions (3D) human reconstruction methods in the past five years were summarized．First，a series of the current state-of-the-art single-image 3D human reconstruction methods were combed from both the perspectives of model representation and computing method． For model representation，the four common representations，including depth image and point cloud representation，parametric body model representation，voxel and semantic voxel representation，and implicit surface function representation，as well as their mutual transformation relationship were presented in detail． For computing method，the proposed algorithms based on the above four representations were deeply described，and their pros and cons were analysed． Subsequently，the publicly available datasets for single-image 3D human reconstruction were introduced，and the quantitative evaluation metrics were presented．Then，the state-of-the-art single-image 3D human reconstruction methods were evaluated and compared quantitively and qualitatively on publicly available datasets． Finally，based on the experimental results，the problems of the existing methods were presented，and future challenges and research directions of single-image 3D human reconstruction were discussed. © 2024 Huazhong University of Science and Technology. All rights reserved.