3D object recovery and stylization with limited supervision
The acquisition of 3D data is often costly and challenging, leading to a scarcity of reliable 3D ground truth for training deep learning models. This thesis focuses on 3D tasks that involve limited supervision, where access to comprehensive training data is constrained. Specifically, this thesis con...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Doctor of Philosophy |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/174068 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | The acquisition of 3D data is often costly and challenging, leading to a scarcity of reliable 3D ground truth for training deep learning models. This thesis focuses on 3D tasks that involve limited supervision, where access to comprehensive training data is constrained. Specifically, this thesis concentrates on three tasks: (1) 3D shape completion, where paired partial-complete shapes for training are severely limited; (2) 3D textured mesh reconstruction, where the training data consists solely of a collection of monocular images; and (3) 3D face toonification, where only 2D pseudo ground truths are available. To address these challenges, I explore the utilization of pre-trained 3D Generative Adversarial Networks (GANs) as valuable sources of geometric and visual priors. By harnessing the rich prior knowledge captured by these pre- trained models, the proposed methods enable accurate and robust completion, reconstruction, and stylization in these challenging 3D scenarios with limited supervision. |
|---|