基于决策图的三维模型无监督聚类算法

Unsupervised clustering algorithm on 3D model library via decision graph

针对三维模型的无监督聚类问题,目前广泛采用基于词袋的方法具有两大缺陷,既无法准确知道聚类的数目,也不能适用于结构复杂(比如呈流形结构)的形状空间.为此,本文采用两大方法加以改进,其一利用有流形聚类功能的决策图方法取代K-means,其二使用核函数更加有效地衡量三维形状之间的差异.在SHREC2010库和SHREC2011库上的大量实验结果表明,两种技巧的有机结合使聚类的精确度和效率得到了显著的提升.

In the unsupervised clustering problem of 3 D models, the Bag-of-Word based approach has been identified with two major drawbacks： being unable to predict exactly the number of clusters, and unable to apply to a complicated shape space（e.g., with a manifold structure）. Therefore, we propose two improvement techniques in this paper. Firstly, the decision graph technique, with a super ability in manifold clustering, is used as a substitute for K-means. Secondly, the difference between shapes is calculated using a combined kernel function, rather than measured in the original signature space. Extensive experimental results on SHREC2010 and SHREC2011 show that the coherent combination of the couple of techniques greatly improves clustering accuracy and efficiency.