Single view image 3D geometry estimation using self-supervised machine learning

Zhou, Hang (2024) Single view image 3D geometry estimation using self-supervised machine learning. Doctoral thesis, University of East Anglia.

Preview

PDF Download (35MB) | Preview

Abstract

Recovering 3D information from 2D RGB images is an essential task for many applications such as autonomous driving, robotics, and augmented reality, etc. Specifically, estimating depth information, which is lost during image formation, is a vital step for downstream tasks. With the development of deep learning, especially supervised learning, more and more researchers exploit this technique to improve depth estimation. However, supervised learning based models’ performance heavily relies on the quality of depth ground truth which is expensive to collect. In contrast to supervised learning methods, based on well-established Structure-from-Motion, self-supervised approaches only require sequential images to train depth estimation models, which transfer a depth regression task to an image reconstruction task.

In this thesis, we focus on improving self-supervised monocular depth estimation. To this end, we propose several approaches: Firstly, we explore temporal geometry consistencies across consecutive frames and propose a depth loss and a pose loss. Secondly, we adopt HRNet and attention mechanism to build a novel representation network architecture DIFFNet, which significantly benefits from higher resolution input images. Thirdly, we propose a two-stage training scheme upon the existing one-stage framework by introducing a second-stage training when a self-distillation loss is optimized at the same time as the photometric loss. All of my works have been published at conferences.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Nicola Veasy
Date Deposited:	11 Jul 2024 09:17
Last Modified:	11 Jul 2024 09:17
URI:	https://ueaeprints.uea.ac.uk/id/eprint/95863
DOI:

Downloads

Actions (login required)

View Item