SVNET: A SINGLE VIEW NETWORK FOR 3D SHAPE RECOGNITION

As the great success of deep learning in the 2D image recognition, applying it to 3D shape recognition has recently received much more attention. In this paper, we propose a Convolutional Neural Network (CNN) framework named Single-view Network (SVNet) for 3D recognition. Unlike multiple-view-based methods that aggregate multiple views into one, SVNet extracts and retains the feature of each view separately. Concretely, there are M fully connected layers in the last layer of SVNet, where M denotes the number of views. SVNet integrates the prediction of each fully connected layer to get the final result, which is similar to voting. In addition, our method does data augment by aligning the 3D shape and using normalized normals as the color of 3D shapes. Compared to the state-of-the-art methods, SVNet achieves better performance in 3D shape classification and retrieval on the benchmark dataset. The implementation of SVNet is available at https://github.com/paopaoer/experiment.git.