基于LL-GPLVM的目标形状建模方法研究

Research on Target Shape Modeling Based on Local Linear Gaussian Process Latent Variable Model

为实现利用二维图像对三维目标外观进行准确建模,提出了一种基于局部线性约束的高斯过程隐含变量模型(LL-GPLVM)的目标形状建模算法.该算法以形状作为目标特征表示,以半球形作为流形拓扑结构约束,用LL-GPLVM来完成目标视角流形学习.目标视角流形是一个概率模型,关于流形结构的先验知识以局部线性约束的方式引入,学习结果为数据与先验知识的平衡结果.对仿真数据的实验结果表明,该视角流形能准确地表示目标形状,建模精度高于现有算法.

A multi-view target shape modeling method based on local linear gaussian process latent variable model （LL-GPLVM） was proposed which could learn a shape manifold from a set of 2D shapes for 3D target modeling. LL-GPLVM was used to learn a view manifold for the target, in which shape was used for feature representation, a hemisphere was used as topology constrain for manifold learning. The prior of the manifold structure was involved in manifold learning through local linear embedding, so the final manifold was the balanced result of data and topology prior, and was a probabilistic model. Experiment on synthetic data demonstrates the advantages of the proposed method over existing techniques.