利用空间结构信息的三维点云模型分类

现有的三维点云模型分类方法未考虑模型本身的空间结构信息,忽略了模型上点与点之间的相互关系.为此,提出一种能够提取模型空间结构信息的转换网络,实现三维点云模型的分类.首先对三维模型采样分组,得到其球形邻域,计算每个邻域内点的浅层特征,同时使用转换网络将邻域的空间结构信息转换为特征权重,并通过特征映射将特征权重和浅层特征输出为具有该邻域空间结构信息的高维特征.然后聚合各个邻域的高维特征得到模型的全局特征,并通过多个尺度逐层迭代输出分类结果.实验结果表明,在ModelNet40上的分类准确率达到92.8%,高于目前的主流算法.

3D Model Classification Using Spatial Structure Information

The existing classification methods of 3D point cloud models do not consider the spatial structure information of the model itself,and have not yet paid enough attention on interactions between points on the models.Therefore,a spatial structure information transformation network,SSITNet,is proposed to complete classification by extracting the spatial structure information of 3D models.Firstly,the 3D model is sampled and grouped to obtain its spherical neighborhood,and low level features of points in each neighborhood are calculated.At the same time,the SSITNet is used to transform the spatial structure information of the neighborhood into feature weights.Through feature mapping,high level features with spatial structure information of the neighborhood are obtained from the low level features and their weights.Then the global features of the model are obtained by aggregating the high level features of each neighborhood and the classification results are achieved by multiple-layer iterations.Experimental results show that the classification accuracy of our algorithm on ModelNet40 is 92.8%,which is better than the current mainstream algorithms.