基于分层学习的三维模型兴趣点提取算法

针对基于学习的三维模型兴趣点提取问题,提出一种兴趣点分层学习的全监督算法.提取三维模型表面所有顶点的特征向量后,将人工标注的兴趣点分为稀疏点和密集点,对于稀疏点使用整个三维模型进行神经网络训练,对于密集点则找出兴趣点分布密集的区域进行单独的神经网络训练;然后对2个神经网络进行特征匹配,得到一个用于三维模型兴趣点提取预测的分类器.测试时,提取新输入的三维模型上所有顶点的特征向量,将其输入到训练好的分类器中进行预测,应用改进的密度峰值聚类算法提取兴趣点.算法采用分层学习的策略,解决了传统算法在模型细节处难以准确提取密集兴趣点的问题.在SHREC’11数据集上的实验结果表明,与传统算法相比,该算法提取兴趣点的准确率更高,出现的遗漏点和错误点更少,对解决越来越精细的三维模型的兴趣点提取问题有较大帮助.

Detecting 3D Points of Interest Using Hierarchical Training Strategy

In this paper,we propose a novel supervised 3D points of interest(POIs)detection algorithm by using hierarchical training strategy.Firstly,the feature vectors of all the vertices of training shape are extracted,and the labeled POIs are divided into the part with sparse points and the part with dense points.Secondly,the two grouped POIs are used to train neural networks.Finally,a 3D shape POIs classifier is obtained by matching the two neural networks with the feature vectors.In the testing process,the feature vectors of all the vertices are extracted and fed to the trained classifier for prediction.An improved density peak clustering algorithm is then used to detect the POIs.Our algorithm adopts the hierarchical training strategy,which can address the issue of accurately detecting the dense POIs in the model with details.The experimental results show that our method detects the POIs with higher accuracy when compared with the traditional algorithms.Both the false positive error and false negative error are greatly reduced by using our hierarchical training method.