Deep Visual Learning with the Aid of Unlabelled Data

Abstract

In this thesis, research is conducted on computer vision problems across several sub-fields, including image classification, point cloud classification and segmentation, with a focus on approaches based on deep neural networks (DNNs). The corpus of research in computer vision has greatly advanced over the past decade thanks largely to DNNs, enabling and facilitating various real-world vision applications. Despite the promising results achieved, the major problem with using DNNs for supervised learning tasks is that they usually demand large amounts of data to train, which sets bottlenecks for applications where the data are difficult or expensive to annotate. To alleviate this data shortage issue, there are recent studies on DNN-based semi-supervised learning and domain adaptation which improve the generalisation of DNNs by exploiting the additional information contained in the unlabelled data during model training. By following semi-supervised learning and domain adaptation paradigms, several computer vision problems are tackled in this thesis. This thesis presents several contributions to deep image and point cloud learning that exploit unlabelled data. First, a comprehensive literature review covering deep learning, point cloud learning, semi-supervised learning, and domain adaptation is conducted, and a thorough introduction to the theoretical backgrounds is provided. Second, a generic theoretical framework for semi-supervised learning is proposed and a semi-supervised learning method for image and point cloud classification called Augmented Distribution Alignment (ADA-Net) is presented. Third, a cross-dataset point cloud classification method called Deep-Shallow Domain Adaptation Network (DSDAN) is proposed which combines the advantages of both DNN and handcrafted features for better point cloud representation while utilising the unlabelled target data for improved cross-dataset generalisation. Fourth, this thesis focuses on the real-world application problem of forest inventory and propose a semi-supervised and domain adaptation framework for tree point cloud semantic segmentation as well as a DNN-based tree parameter estimation. Finally, this thesis casts insight into the future research directions of semi-supervised learning and domain adaptation for image and point cloud learning.