一种三维模型多层级视点描述符

A multi-level viewpoint descriptor for 3D model

3D形状识别是计算机视觉领域非常重要的问题之一,三维模型的描述符对不同类别模型的区分度影响着3D形状识别的精度。基于视图的描述符应用非常广泛,然而多数方法主要研究算法,以不断提升公共数据集的识别精度,却很少深入分析在细粒度分类数据上的视图选择对区分能力的影响。该文提出一种新的多层级视点描述符,该描述符以一组视点组成,为3D模型提供了一组最优的二维视图选择,并设计实验对比在细粒度分类数据上的识别精度。首先,我们建立均匀的球状观察视点模型,以具备直观语义信息的上、前、右等视点为初始视点。其次,分区域从模型投影边缘的倒角距离比较中得出最优视点,作为下一层的种子点,不断细分获取不同层级的视点组合得到多层级视点描述符,使用测地距离来衡量模型间的相似度。最后,使用视点间的测地距离来计算同类别及不同类别间的相似度,验证多层级视点描述符的区分能力,选择更加精细分类的三维模型数据以验证识别能力,在多视图卷积神经网络上进行对比。应用该文描述符提供的视图选择在训练数据无论是否进行随机水平旋转,都有非常明显的识别精度提升。

3D shape recognition is one of core problems in computer vision.3D descriptors have distinguishable capacity to different categories of models,which directly affects the accuracy of 3D object recognition.View based descriptors are widely used.However,most methods mainly study algorithms to improve the recognition accuracy of common data sets,but seldom analyze the impact of view selection on the discrimination ability of fine-grained classification data.In this paper,we propose a novel view-based multilevel viewpoint descriptor(MLVD)composing a set of viewpoints,which provides a set of best view selection.We first establish dense uniform viewpoints on the view sphere,and then the semantic intuitive viewpoints such as like top,left and right,are selected as the initial viewpoints.Afterwards,each best viewpoint is obtained as the seed point of the next layer from the comparison of silhouette images using chamfer distance,and further subdivision forms different level 3D descriptors.Finally,we calculate similarities between the 3D shape model from the same class and different class using geodesic distance of viewpoint for verifying the discrimination capacity of MLVD.In addition,3D model recognition experiment is carried out on multi-view convolutional neural networks(MVCNN).Using the view selection provided by the descriptors in this paper,no matter whether the training data is rotated randomly or not,the recognition accuracy is improved obviously.