Form Generation with Embedded 3D Modeling and Spatial Deep Learning

This work addresses the challenge of reconstructing 3D models from either single or multi-view images by leveraging implicit representations and geometric deep learning methods. In the area of Computer Vision, the challenge of 3D reconstruction has attracted a lot of interest over an extended period due to its diverse set of practical applications, such as robotics(Li et al., 2021), augmented & virtual reality (Esperto et al., 2021), and medical imaging (Molaei et al., 2023). However, traditional methods for 3D reconstruction suffer from limitations (Aharchi and Ait Kbir, 2020) such as high computational complexity and sensitivity to noise and errors. To address these limitations, several approaches has been "proposed" for 3D reconstruction using implicit representations, which yields a compact & continuous representation of 3D shapes. This chapter began with a clear explanation of the research problem, followed by an exploration of its underlying motivations. The concept of ill-posedness(Wang et al., 2018) in 3D reconstruction was subsequently elucidated to provide a heightened sense of clarity. To reinforce the foundational knowledge, an exhaustive examination of relevant literature was undertaken, covering the domains of 3D reconstruction, implicit representations, and deep learning within the subject area of computer vision. Advancing further, an innovative set of approaches for single-view 3D reconstruction using implicit representations was introduced. These methods were carefully crafted to tackle the complexities of the specific challenge. To substantiate the potential of these techniques, a series of experiments were conducted across different datasets.