针对着装图像的参数化人体重建

Reconstructing parametric human body from dressed human photo is very important to virtual fitting systems. Its main difficulties include: There is no method to accurately describe the transformation from naked model to dressed model; When optimizing the parameters， the distance from the human body to the target data is expensive to calculate. About the first difficulty, we try to generate a training database composed of parametric human body and its corresponding dressed image， and learn from the training data to obtain the naked body of an input dressed human photo. We need large number of samples to use statistic learning. We plan to generate the corresponding dressed body by cloth simulation. About the second difficulty, we try to mark some key points on the 3D human body and use these key points to bridge the dressed photo with the 3D body mesh. To avoid the expensive calculation, we obtain the key points corresponding to dressed photo by statistic learning and use these key points to optimizing the parameters of human body. We will also research on a more intrinsic parametric model suitable for statistic learning framework. If this proposal can be carried out smoothly, it can not only find a way to reconstruct human body from dressed human photo, but also provide reference for a series of similar problems such as finding 3D skeleton from human photo.

从着装图像快速重建参数化人体模型对虚拟试衣系统有重要意义，其难点在于：一是无法准确地描述着装模型到三维人体模型的变换；二是在优化人体参数使人体模型拟合目标数据时两者的距离计算非常耗时。针对第一个难点，本项目拟构建由参数化人体模型及该模型相应的着装图像构成的训练库，通过对训练库的学习得到与输入的着装图像相对应的三维人体，采用学习的方法需要大量样本，我们拟采用基于物理的布料仿真来模拟三维人体模型穿上衣服的效果，以获取训练数据；针对第二个难点，我们拟通过在三维人体模型上标记稀疏标志点，以稀疏标志点为桥梁，通过统计学习获得着装图像对应的标志点，然后用这些标志点构建目标函数优化人体参数，从而避免耗时的距离计算。同时我们还将研究更直观且易于使用统计学习框架的人体参数化方法。本项目若能完成，不仅能解决从着装图像重建人体这一难点问题，同时对从图像提取三维人体骨架等一系列问题都有很大借鉴意义。

从着装图像快速重建参数化人体模型对虚拟试衣系统具有重要意义，其难点在于：一是无法准确地描述着装模型到三维人体模型的变换；二是在优化人体参数使人体模型拟合目标数据时两者的距离计算非常耗时。针对第一个难点，本项目基于布料仿真技术构建了首个含大量三维未着装及着装人体对的数据库，通过对训练库的学习得到与输入的着装图像相对应的三维人体。针对第二个难点，我们设计了一组与人体身材尺寸密切相关的标志点，以稀疏标志点为桥梁，通过统计学习获得着装图像对应的标志点，然后用这些标志点构建目标函数优化人体参数，从而避免耗时的距离计算。同时我们研究从着装图像直接回归人体参数，进一步提升了人体重建效率。本项目针对着装图像的参数化人体重建工作在重建效率及方便性方面的提升，促进了面向普通网购用户的虚拟试衣系统的普及，重建精度也满足试衣需求。此外，本项目提出的语义身材模型弥补了现有参数化人体建模在直观刻画人体身材方面的不足，基于GPU加速的布料仿真方法在仿真效果及效率方面的提升对虚拟试衣、游戏、电影等相关行业具有重大意义。