View-Based 3D Model Retrieval via Convolutional Neural Networks

In this paper, we propose a multi-view fusion 3D model retrieval using convolutional neural network to solve the problem of the local perception in feature descriptor. By view pooling, we combine information from multiple views of a 3D model to eliminate the position correlation caused by the viewing angle of camera. In addition, integrating pre-processed RGB view-feature with Binary view-feature in the same model is used to generate a single model descriptor. Experiments on ETH dataset demonstrate the superiority of the proposed method.