Multi-view Fusion with Deep Learning for 3D Shape Classification

3D model retrieval is all along a difficult and hotspot in computer vision. Recently, the view-based methods make use of the advanced convolutional neural networks which achieve excellent results. However, the structural information of the 3D model was destroyed by the projection and the relevance of multiple perspectives was not considered. In order to resolve this problem, this manuscript analyzes the process of human observation of 3D models and imitates the process of human recognition of 3D models through a combination of convolutional neural networks and recurrent neural networks. Our approach can convert the relevance between perspectives into the dependency relationship between structured recurrent neural networks and restore the structural information of the 3D model. We conducted experiments on the ModelNet40 database, it shows strong performance on par or even better than state-of-the-art results.