Bachelor Thesis BCLR-2022-74

BibliographyManea, Radu: Text-to-CAD-Model-Part Synthesis: A Feasibility Study on Adapting Machine Learning Techniques from Image Synthesis.
University of Stuttgart, Faculty of Computer Science, Electrical Engineering, and Information Technology, Bachelor Thesis No. 74 (2022).
70 pages, english.

The process of generating data based on a given description is rapidly becoming one of the most significant areas of machine learning. Image synthesis and CAD model synthesis are two common applications in this domain. There are a variety of uses for software that can generate images or 3D models based on a prompt. Not only are such systems capable of generating useful data, but they can also be used for labeling data and retrieving unlabeled data. This thesis focuses on text-to-CAD-model synthesis, which is the creation of 3D models from captions. The field of 3D synthesis has not received as much attention as the image synthesis domain, which is why the focus of this work is on 3D models. Studies that have been done in the past on this subject have primarily concentrated their attention on a relatively small subset of objects, such as tables and chairs. This does not correspond to the diverse array of objects that such a system ought to be able to generate. In addition, the proposed solutions do not place an emphasis on efficiency, which leads to the creation of applications that require a significant amount of resources and are difficult to instruct. The purpose of this feasibility study is to determine whether or not it is possible to perform text-to-shape synthesis in an effective manner by utilizing a variety of methods from other domains, such as the text-to-image domain, amongst others. The purpose of this experiment is to determine whether or not a system can be devised that is capable of generating a wide range of different objects while maintaining a high level of efficiency. We have obtained satisfactory results, demonstrating that it is possible to create CAD models from text for 55 classes of objects. In addition to this, a significant amount of development has taken place in terms of the effectiveness of the method. We have been successful in demonstrating that a latent-space generative model can work for the synthesis of models while simultaneously maintaining its efficiency. In addition to describing which techniques delivered promising results, we also discuss the issues and difficulties that arose throughout the course of the experiments. Our research lays the groundwork for an effective approach to the problem of 3D model synthesis.

Department(s)University of Stuttgart, Institute of Software Technology, Empirical Software Engineering
Superviser(s)Wagner, Prof. Stefan; Schoenhof, Raoul; Bogner, Dr. Justus
Entry dateNovember 29, 2022
   Publ. Computer Science