时序深度玻尔兹曼机在Mocap数据修复中的应用

Application of Temporal Deep Boltzmann Machine in Recovering Motion Capture Data

本文提出了一种利用时序深度玻尔兹曼机网络估计动作捕获中缺少的标记数据的方法,从而得到流畅的动作数据。本方法即使对于多帧多人体部位的全部标记数据缺失问题仍能得到精确问题的结果。另外,该方法训练的深度网络参数可以对应不同类型的人体动作。因此,只要新来的动作序列中相似的运动数据收集在网络训练集中,本方法即可对此序列进行缺失数据修复,这比传统方法更容易设置训练集。

In this paper, we propose a method to estimate the missing marker data in motion capture by using temporal deep Boltzmann machine network so as to obtain smooth motion data. This method can get accurate results even if all the labeled data of multi frame and multi body parts are missing. In addition, the deep network parameters trained can correspond to different types of human actions. Therefore, as long as the similar motion data in the new action sequence are collected in the network training set, this method can repair the missing data of this sequence, which is easier to set up the training set than traditional methods.