Learning to complete and reconstruct 3D models with sparse representations

Wang, Tao 2023. Learning to complete and reconstruct 3D models with sparse representations. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (PhD thesis) - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (9MB) | Preview
	PDF (Cardiff University Electronic Thesis and Dissertation Form) - Supplemental Material Restricted to Repository staff only Download (312kB)

Abstract

With the development of acquisition techniques, the input source for computer vision has become more diverse. Traditional RGB colour channels could be captured with auxiliary depth information, and the research field of computer vision has been expanded to three dimensions. Information extracted from 3D shapes is not only for recognition tasks but also beneficial to geometric enhancement and new content generation. Therefore, 3D vision has a strong bond with traditional computer graphics. For example, a surface approximation module could be embedded into a pipeline by taking encoder-extracted features of a noisy point cloud as input to perform 3D surface reconstruction. Although there are a few attempts on this task, better detail recovery and robustness to noise are still in demand. Besides, data acquisition is the first step in a typical geometric processing pipeline. The enhancement of defective raw input is crucial for downstream processing. Both enhancement of captured 3D data and synthesis of new content rely on effective visual perception ability of learning-based models. This thesis offers a solution to the aforementioned challenges from the angle of the spatial organisation of data. As 3D shapes are 2D surfaces embedded in the 3D space, various effective data structures for efficient processing have been developed before. Using a sparsified representation has been attempted in this thesis to improve the result of 3D completion and 3D reconstruction. Specifically, this thesis proposes a pipeline using convolutions on sparse tensors with a carefully designed embedding of points’ distribution to bring prominent results on indoor scene completion. Additionally, the sparse representation of the noisy point cloud’s latent space facilitates accurate feature interpolation, thus regressing an implicit function for 3D mesh reconstruction. Lastly, sparse voxel grids have been utilised to reconstruct radiance fields from single-view image input.

Item Type:	Thesis (PhD)
Date Type:	Acceptance
Status:	Unpublished
Schools:	Computer Science & Informatics Engineering
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Funders:	China Scholarship Council
Date of First Compliant Deposit:	5 June 2024
Date of Acceptance:	December 2023
Last Modified:	06 Jun 2024 13:21
URI:	https://orca.cardiff.ac.uk/id/eprint/169491

Actions (repository staff only)

Edit Item