| Abstract [eng] |
This study explores the use of deep learning techniques with three-dimensional representational forms. The literature review outlines common methods for deep learning from various 3D data formats as well as the benefits and drawbacks of 3D representational forms. The evaluation and comparison of the 3D-GAN generative adversarial neural network’s generation outputs are the primary objectives of the study. The research involves measuring the resemblance and realism of 3D-GAN derived 3D models by employing a variety of similarity metrics to examine the similarity between the generated 3D voxel grids and the ground truth. The study also examines the impact of 3D voxel grid density and resolution on the estimate of similarity measures to better understand the elements impacting 3D voxel grid similarity scores. Overall, the study tackles recurring challenges in the evaluation of quality and realism of 3D-GAN produced models. |
