3D object understanding with 3D Convolutional Neural Networks

Feature engineering plays an important role in object understanding. Expressive discriminative features can guarantee the success of object understanding tasks. With remarkable ability of data abstraction, deep hierarchy architecture has the potential to represent objects. For 3D objects with multiple views, the existing deep learning methods can not handle all the views with high quality. In this paper, we propose a 3D convolutional neural network, a deep hierarchy model which has a similar structure with convolutional neural network. We employ stochastic gradient descent (SGD) method to pretrain the convolutional layer, and then a back-propagation method is proposed to fine-tune the whole network. Finally, we use the result of the two phases for 3D object retrieval. The proposed method is shown to out-perform the state-of-the-art approaches by experiments conducted on publicly available 3D object datasets.