Use this URL to cite or link to this record in EThOS:
Title: Feature driven learning techniques for 3D shape segmentation
Author: George, David
Awarding Body: Swansea University
Current Institution: Swansea University
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Segmentation is a fundamental problem in 3D shape analysis and machine learning. The ability to partition a 3D shape into meaningful or functional parts is a vital ingredient of many downstream applications like shape matching, classification and retrieval. Early segmentation methods were based on approaches like fitting primitive shapes to parts or extracting segmentations from feature points. However, such methods had limited success on shapes with more complex geometry. Observing this, research began using geometric features to aid the segmentation, as certain features (e.g. Shape Diameter Function (SDF)) are less sensitive to complex geometry. This trend was also incorporated in the shift to set-wide segmentations, called co-segmentation, which provides a consistent segmentation throughout a shape dataset, meaning similar parts have the same segment identifier. The idea of co-segmentation is that a set of same class shapes (i.e. chairs) contain more information about the class than a single shape would, which could lead to an overall improvement to the segmentation of the individual shapes. Over the past decade many different approaches of co-segmentation have been explored covering supervised, unsupervised and even user-driven active learning. In each of the areas, there has been widely adopted use of geometric features to aid proposed segmentation algorithms, with each method typically using different combinations of features. The aim of this thesis is to explore these different areas of 3D shape segmentation, perform an analysis of the effectiveness of geometric features in these areas and tackle core issues that currently exist in the literature. Initially, we explore the area of unsupervised segmentation, specifically looking at co-segmentation, and perform an analysis of several different geometric features. Our analysis is intended to compare the different features in a single unsupervised pipeline to evaluate their usefulness and determine their strengths and weaknesses. Our analysis also includes several features that have not yet been explored in unsupervised segmentation but have been shown effective in other areas. Later, with the ever increasing popularity of deep learning, we explore the area of super-vised segmentation and investigate the current state of Neural Network (NN) driven techniques. We specifically observe limitations in the current state-of-the-art and propose a novel Convolu-tional Neural Network (CNN) based method which operates on multi-scale geometric features to gain more information about the shapes being segmented. We also perform an evaluation of several different supervised segmentation methods using the same input features, but with vary-ing complexity of model design. This is intended to see if the more complex models provide a significant performance increase. Lastly, we explore the user-driven area of active learning, to tackle the large amounts of inconsistencies in current ground truth segmentation, which are vital for most segmentation methods. Active learning has been used to great effect for ground truth generation in the past, so we present a novel active learning framework using deep learning and geometric features to assist the user in co-segmentation of a dataset. Our method emphasises segmentation accu-racy while minimising user effort, providing an interactive visualisation for co-segmentation analysis and the application of automated optimisation tools. In this thesis we explore the effectiveness of different geometric features across varying segmentation tasks, providing an in-depth analysis and comparison of state-of-the-art methods.
Supervisor: Xie, Xianghua ; Tam, Gary K. L. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral