基于高斯过程的三维模型语义分类和检索

针对三维模型检索系统提高准确率、减少几何特征和人类语义丰富性之间的“语义鸿沟”等问题, 提出一种基于高斯过程的语义分类和检索新方法.该方法采用一种统计2个采样点相对质心向量夹角的AC2直方图新特征,与形状分布的D2特征组合成低层特征，使用高斯过程进行三维模型语义分类的监督学习,计算测试模型的语义类概率预测分布,建立低层特征和查询概念之间的联系；使用语义距离和不相似度计算方法进行检索排序.实验结果表明：与已有的某些监督学习的方法相比,多类的测试模型进行语义分类的准确率明显得到提升,检索中能体现语义概念,检索性能也得到提高.

3D model semantic classification and retrieval with Gaussian processes

A novel 3D model retrieval and semantic classification method using Gaussian processes was proposed to improve the performance of 3D model retrieval systems, and reduce the ‘semantic gap’ between the shape features and the richness of human semantics. A new type of feature named AC2 using histogram of angle between the centroid and pairs of random points was proposed, which combined D2 of shape distribute as low-level feature. The Gaussian processes were used for 3D model semantic classification as supervised learning, and the predictive distribution of the semantic class probability was computed for associating low-level features with query concepts. The method ranked models by dissimilarity measure incorporating the semantic distance and the shape feature distance. Experimental results showed that the multi-class 3D model classification accuracy using the proposed method is significantly higher than those of other supervised learning methods, and the retrieval can capture the query model’s semantics, so the performance is improved.