经典人体三维重建通过主动式深度图融合技术或多视点三维重建方法实现，无法突破重建精度和效率的固有矛盾；而单视图人体三维重建通过任意单张图片恢复人体三维外形，成为了影视制作、智慧城市、辅助医疗等领域迫切需求的前沿技术。当前主流的单视图人体重建方法普遍基于深度学习框架，面临着重建精度低、泛化能力差问题。本项目旨在通过设计高效的人体外形表示方式，并建立面向动态人体的自监督式学习框架，研究深度神经网络对三维外形预测的内在规律，解决数据缺乏和模型训练之间的矛盾，为单视图三维重建提供新原理和新技术，并建立大规模、高精度、衣着外貌多样化的三维人体数据库。申请人在人体三维重建领域积累了丰富的研究基础，构建了高精度多视点计算摄像平台，应用于《爵迹》等影视内容的特效制作，发表包括CVPR、IEEE T-CSVT在内的论文9篇，申请专利7项。

The classical 3D reconstruction of human is realized by active depth-map fusion technology or multi-view 3D reconstruction method, which cannot break the inherent contradiction between reconstruction accuracy and efficiency. The single-view 3D human reconstruction method can recover human shape by any single picture, which has become an urgent need in the field of film and television production, smart city, auxiliary medicine and other cutting-edge technology. The current mainstream single-view body reconstruction methods are generally based on the deep learning framework, and are facing the problems of low reconstruction accuracy and poor generalization ability. This project aims to design efficient human shape representation, and establish the dynamic human oriented self-supervised learning framework, study the inherent law of deep neural network to predict the 3D shape, solve the contradiction between the lack of data and model training, provide new theory and new technology on single-view 3D reconstruction, and establish the large-scale detailed 3D human dataset with various cloth and appearance. The applicant has accumulated a rich research foundation in the field of 3D reconstruction of huma, built a high-precision multi-view computing photography platform, applied it to the special effects production of film and television content such as ‘the lord’, published 9 papers including CVPR and IEEE T-CSVT, and applied for 7 patents.