3d Shape Recognition And Cross-domain Retrieval Based On Multi-view Data

With the rapid growth of the computer technology and popularization of the internet,it's more and more convenient for people to get access to all kinds of information.Meanwhile,the amount of information has soured with more and more types.Compared with pictures and videos,3D models are relatively new but critical.Nowadays,3D models are applied widely in many areas such as product design industry,films and animation,electronic commerce,historic preservation,urban planning,VR and so on.With the dropping of the cost of obtaining 3D models and the increasing demand for 3D models applications,3D models are produced in quantity,a portion of which are open sourced for academic and other usages.How to make analysis for 3D models in an effective way is a hot topic.Many researchers propose methods towards 3D model recognition.Among these works,analysis based on point cloud data and multi-view data has achieved quite good performance and is becoming the mainstream approaches.In this paper,we focus on the multi-view data of 3D models recognition and cross domain retrieval and propose two methods.1.Considering that the view images has spatial continuity,we cannot ignore the important information.For recognition task,we propose a two stream network with progressive feature guide learning(PGNet)based on multi-view information of 3D shape,which includes two sub-networks including MVCNN and LSTM to extract the view information and fuse together to a better feature.2.For retrieval task,we propose a novel deep correlated joint networks(DCJN)method to bridge the gap between 2D images and 3D models,which jointly trains two distinct deep neural networks,which learns two deep nonlinear transformations to extract features from both modals into the same feature space.The final experimental results demonstrate the superiority of our proposed methods over state-of-the-art methods.