Abstract: With the progress of data acquisition equipment and modeling technology, how to efficiently analyze and retrieve 3D shapes has become a hot topic in the field of geometric processing. Currently, many works are focused on the classification of models, almost only consider single label. When dealing with the multi-label problem, the correlation between tags and samples is ignored ,as a result, a lot of computation time is spent. To address this problem, a 3d model labeling method based on multi-label propagation is proposed. The core of this approach is to explore the potential of multi-label correlation in the whole sample space by using the label and sample correlation. In particular, a small set of multi-labeled samples are given, it iterates dynamic metric fusion with label information and the label information of these samples are propagated to the unlabeled samples. The process of propagation depends on similarity matrices between samples and the relationship between labels and samples is taken into account. The labeling results are gotten until the iteration finish. Experimental results on the PSB dataset demonstrates the high accuracy of the approach with a few interactions.

Key words: 3D shape classification, multi-label, label propagation

摘要： 随着数据采集设备与建模技术的进步，如何高效地对三维模型进行分析与检索，成为目前几何处理领域的研究热点。当前，有许多工作都集中在模型的分类上，但是大多仅能处理单一标签。在处理多标签问题时，不仅耗费大量时间还忽略了标签之间和样本之间的关联关系。针对该问题，提出了采用多标签传播的三维模型标注方法。其核心在于利用标签相关性与样本之间的关联关系探索到整个样本空间的多标签标注潜力。具体来说，给定一小部分样本的多标签信息，再将多标签信息通过这些标注样本传播到空间中无标注的样本之上。传播的过程主要依靠迭代融合标签信息与动态度量，充分考虑了标签之间与样本之间的关联关系，最终得到整个空间的标注结果。在一些三维模型的标准数据集上（如普林斯顿形状标准模型数据库）进行实验测试，结果证明，只需要少量的交互就能快速地得到较为精确的结果。

关键词: 三维模型分类, 多标签, 标签传播