3D GANs For Emirati Faces
Bin Kuwair, Leena Ali Faisal Salem
Bin Kuwair, Leena Ali Faisal Salem
Supervisor
Department
Computer Vision
Embargo End Date
Type
Thesis
Date
2024
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This paper evaluates the EG3D technology developed by Nvidia, employing our resources to experiment with and modify this advanced project. EG3D was initially trained on eight Tesla V100 GPUs consisting of seven days of training with a batch size of 642, followed by 1.5 days of fine-tuning with a batch size of 1282. However, when we applied this method using a single GPU and a 4 GPU setup, the results were significantly inferior compared to the original high-resource configuration. This outcome reveals a critical scalability issue, indicating that although Nvidia has made the source code available, the methods are tailored for high-end environments and falter on less powerful hardware. This disparity underscores the technology's limited accessibility for research across varied hardware capabilities.
Despite these challenges, our study, titled "3D GANs for Emirati Faces," successfully introduces new data processing techniques that enhance results. We adapted the Multi-task Cascaded Convolutional Networks (MTCNN) to detect multiple facial landmarks and integrated it with human pose estimation methodologies, enabling us to capture not only the faces but also the shoulders of the subjects. This enhancement allows for a more comprehensive display of the subjects' attire and cultural backgrounds in three dimensions. Furthermore, we developed a new sample dataset of Emirati individuals specifically for this project, enabling rigorous testing and fine-tuning of our approaches.
Citation
L. A. Bin Kuwair, "3D GANs For Emirati Faces", M.S. Thesis, Computer Vision, MBZUAI, Abu Dhabi, UAE, 2024