An Adversarial Framework for Deep 3D Target Template Generation

This paper presents a framework for the generation of 3D models. This is an important problem for many reasons. For example, 3D models are important for systems that are involved in target recognition. These systems use 3D models to train up accuracy on identifying real world object. Traditional means of gathering 3D models have limitations that the generation of 3D models can help overcome. The framework uses a novel generative adversarial network (GAN) that learns latent representations of two dimensional views of a model to bootstrap the network’s ability to learn to generate three dimensional objects. The novel architecture is evaluated using two different types of evaluation. The two dimensional views are evaluated using a combination of an Inception Score and Hausdorff Distance, compared against the two dimensional views of the real 3D models used in training. The three dimensional object are evaluated using the Hausdorff Distance compared against the real 3D models used in training. Experimental results demonstrate that the novel generative adversarial network that is being proposed generated realistic looking models faster, and with higher fidelity than a basic 3D generative adversarial network would produce with the same training structure. The thesis illustrates the promise of GAN bootstrapping with two dimensional perspective codes to create higher fidelity three dimensional models