High-quality Shape Synthesis with User-guided Deep Neural Networks

点击翻译按钮获取中文摘要

The long-term goal of my research program is to develop methods for facilitating the creation of computer graphics content, especially 3D shapes. 3D content is important in a variety of applications, such as computer games, movies and animation, and architectural simulations. In the next period of funding, I propose to develop computational methods for synthesizing 3D shapes with deep neural networks (DNNs), where the synthesis can be controlled by users and the generated shapes are of high-quality. These methods can then be applied to facilitate the generation of different types of 3D content. Content creation is challenging since high-quality 3D shapes are commonly created by explicitly modeling the geometry of the shapes. This process involves the use of complex user interfaces by skilled artists in a time-consuming process requiring substantial training. Thus, during the last two decades, computer graphics research has also proposed approaches for facilitating the modeling of 3D shapes by non-expert users, such as parametric models and sketching interfaces. However, many of these methods either still require sufficient artistic skills from the users, or require considerable manual work for pre-processing of the data. In recent years, methods for synthesizing 3D content based on machine learning have sparked much interest, especially methods based on deep neural networks (DNNs), since DNNs offer several advantages over traditional methods, such as easier training data preparation, no need to handcraft feature extraction methods, high generalization capabilities, and generative models that allow users to synthesize new data resembling the training data. However, DNNs currently have certain limitations that prevent them from being easily used for shape modeling. In this context, the goal of my research program for the next period of funding is to investigate solutions for improving the analysis and synthesis of 3D shapes with DNNs. In more detail: (1) We propose to develop methods for enabling more direct user control in the generation of 3D shapes with DNNs, so that users are able to design shapes according to their goals; (2) We propose to use representations that allow to generate higher-quality shapes compared to the state-of-the-art, such as representations based on procedural models that generate editable shapes with low complexity; (3) We propose to use procedural models to generate synthetic data, which can be used for training shape analysis DNNs with less manually-prepared data, or for evaluating DNN-based methods in controlled settings. The significance of the proposed work is that the developed solutions will enable the guided synthesis of high-quality shapes with less manual work. This will result in new software tools that will improve the current practices in industries that require shape modeling, allowing users of this technology to create a diversity and volume of content never seen before with reduced costs.