结合Fisher编码和距离学习的非刚体三维检索方法

摘要

In order to solve the problem of insufficient representation of the non rigid 3D feature descriptor, we propose a new method of non rigid 3D model retrieval based on Fisher Vector coding and distance learning. Firstly, we extract the local descriptor of 3D model to train the dictionary by using Gaussian mix-ture model; then, the local descriptor and the dictionary centers are used as inputs to learn the new feature codes by Fisher Vector encoding method; finally, we map the feature codes with distance learning aiming to construct an efficient global feature with a small inter-class margin and a large intra-class variance, which is used for non rigid 3D model retrieval. We validate our method on the open data sets SHREC10 and SHREC11. The results show that the method achieves a higher accuracy than the traditional method.%针对目前非刚体三维特征描述子表现力不足的问题, 提出一种结合 Fisher 编码和距离学习构建全局特征描述子的非刚体三维模型检索方法. 首先提取三维模型的局部特征描述子, 利用混合高斯模型对特征集合进行无监督字典学习; 然后将局部描述子和字典中心作为输入, 通过 Fisher 编码方法得到新的特征编码; 最后采用距离学习对特征编码进行空间映射, 重构得到类内距离小类间距离大的全局特征描述子, 用于非刚体三维模型检索. 在公开数据集SHREC10和SHREC11上进行实验的结果表明, 相比于传统方法, 该方法在检索精度上有显著提高.