Multi-layers CNNs for 3D Model Retrieval

Due to the rapid development of 3D capturing scanners and better visual process techniques, there is a huge increase of 3D models being uploaded and captured by users. 3D model retrieval has become a hot topic in computer vision. State-of-the-art methods leverage CNNs to solve this problem. But existing CNN architectures and approaches are unable to fully exploit the information of 3D representations. In order to improve the performance of 3D object retrieval algorithms, we proposed a multi-layers CNNs (MLCNN) structure for 3D model representation. First, we combine the 12 rendered views of a 3D object into one representative view, which becomes the actual input. Second, in order to save the global and local information for each 3D model, we aggregate every convolutional layer’s feature into a multi-layers descriptor after a simple PCA compression. Finally, the Euclidean metric is leveraged to compute the similarity between two different 3D models to complete the retrieval problem. The final comparing experiments and corresponding experimental results demonstrate the superiority of our approach.