基于联合特征映射的端到端三维模型草图检索

Joint Feature Mapping for End-to-End Sketch-Based 3D Model Retrieval

在草图-三维模型检索任务中,草图具有类内多样性,三维模型具有复杂性,且草图-三维模型之间存在巨大的域间的差异性,这些特点的相互作用使得基于草图的三维模型检索任务变得特别困难.针对这一问题,提出一种基于联合特征映射的端到端三维模型草图检索框架.首先将三维模型转化为一组二维视图,建立跨域数据的共享数据空间;然后通过网络权值共享,建立端到端的三元度量学习网络,实现跨域数据草图和视图的联合特征映射;最后基于联合特征分布,提出4种草图-三维模型相似评价算法来实现草图-三维模型的检索.在大型公共数据集SHREC2013和SHREC2014上的检索精度分别为81.8%和75.6%,比现有算法在7项检索指标PR曲线,NN,FT,ST,E,DCG和MAP上都有所提升,检索性能突出.

Sketch based 3D model retrieval has the characteristics including diversity of intra-class sketches,complexity of the 3D models,and the huge inter-domain differences between the sketches and 3D models.The interaction of these characteristics makes the sketch-based 3D model retrieval task becomes particularly difficult.To solve the problem,an end-to-end sketch-3D model retrieval framework based on joint feature mapping is proposed.Firstly,the 3D model is transformed into a set of 2D views to establish the shared data space of cross-domain data.Then,through network weight sharing,the end-to-end triplet metric learning network is established,and the joint feature mapping of the cross-domain data,sketches and views,is realized.Finally,based on the joint feature distribution,four kinds of similarity evaluation algorithms between sketches and 3D models are proposed to realize sketch based 3D model retrieval.The retrieval precisions in the large public data sets SHREC2013 and SHREC2014 are 81.8%and 75.6%,respectively,and have demonstrated that the algorithm in this paper is better than the state-of-the art methods in seven indexes of PR curve,NN,FT,ST,E,DCG and MAP.