AN ABSTRACT OF THE THESIS OF Xu Xu for the degree of Master of Science in Computer Science presented on June 13, 2017. Title: Classifying and Synthesizing 3D Shapes of Objects using Deep Neural Networks Abstract approved:

approved: Sinisa Todorovic Reasoning about 3D shape of objects is important for successful computer vision applications in robotics, 3D rendering and modeling. In this thesis, we address two problems – First, given an image, we generate 3D shape of the foreground object that appears in the image. Second, we predict the class label of the input 3D object shape. Recent work uses convolutional neural networks (CNNs) for these problems. However, their training is difficult, since it requires a large amount of training data. Also, in 3D shape generation, existing approaches can not generate realistic 3D shapes and have limited variations in generated 3D shapes. This thesis addresses these issues. We present two novel approaches, one for each problem. First, for 3D classification, we formulate CNN learning as a beam search aimed at identifying an optimal CNN architecture as well as estimating parameters of such an optimal CNN. This model pursuit approach is evaluated on 3D ShapeNet dataset. Second, we introduce a 3D-VAE-GAN (3D Variational AutoEncoder Generative Adversarial Network) model to synthesize high-quality 3D objects from 2D images in ObjectNet3D dataset. Our experimental evaluation shows that our new CNN learning achieves the state-of-the-art results with better modeling efficiency, i.e., with fewer parameters which are much easier to train. Also, in 3D shape synthesis, we achieve larger variability in shapes. c ©Copyright by Xu Xu June 13, 2017 All Rights Reserved Classifying and Synthesizing 3D Shapes of Objects using Deep Neural Networks