基于内外视图的三维模型普适性检索与分类算法
Internal-External View Based for Universal 3D Model Retrieval and Classification
-
摘要: 目前, 基于视图的三维模型检索与分类研究中缺少三维模型内部结构特征, 采用大量外部视图实现非刚性与刚性三维模型的普适性识别. 针对此问题, 提出一种基于内外视图的三维模型检索与分类算法. 首先提出一种内部视图提取模块, 获取内部结构特征; 然后基于正四面体投影提出外部视图提取模块, 获取外部形状特征; 最后采用普适性特征提取网络, 对内外视图进行特征提取与融合学习, 并应用于三维模型检索与分类任务. 在刚性三维模型数据集ModelNet40, 非刚性三维模型数据集SHREC15上的检索平均精度分别达到93.3%和98.9%, 分类整体准确率分别达到94.8%和99.4%. 此外, 在三维模型检索数据集ShapeNet Core55上, 微平均和宏平均的准确率、召回率等系列评价指标表现优异. 表明所提算法能够在少量视图的情况下, 获取显著辨别力的普适性特征.Abstract: The view-based 3D model retrieval and classification algorithms lack the internal structural features of 3D models and require many views to achieve universal recognition of both non-rigid and rigid 3D models. A 3D model retrieval and classification algorithm based on internal and external views is proposed to address this issue. First, an internal view extraction module is proposed to capture internal structural features; then, an external view extraction module is introduced based on tetrahedral projection to obtain external shape features. Finally, a universal feature extraction network is used for feature extraction and fusion learning of internal and external views, applying this to 3D model retrieval and classification tasks. In the rigid 3D model dataset ModelNet40 and the non-rigid dataset SHREC15, the mean average precision reached 93.3% and 98.9%, respectively, with overall accuracies of 94.8% and 99.4%. Additionally, in the ShapeNet Core55, the micro-average and macro-average accuracy and recall metrics demonstrated excellent performance. This indicates that the proposed algorithm captures significant discriminative and universal features with limited views.