Data Efficient Deep Learning Algorithms for Video Processing Systems

Gorpincenko, Artjoms (2022) Data Efficient Deep Learning Algorithms for Video Processing Systems. Doctoral thesis, University of East Anglia.

Preview

PDF Download (34MB) | Preview

Abstract

Over the last decade, deep neural networks have grown in popularity and became the standard approach for the majority of computer vision tasks. Their ability to deliver state of the art results without involving carefully designed hand-crafted features is unmatched, however, that strong performance comes at the cost of requiring large amounts of labelled training data. For many modern applications, collecting and annotating new data can often be time-consuming and expensive, if not impossible altogether.

In this thesis, we focus on scenarios where the cost of acquiring image and video samples and ground truths is the main barrier. We explore different ways to artificially expand the available training sets and propose methods which aim to boost the performance and robustness of neural networks in such situations.

We first visit the field of generative adversarial models and utilise one to mimic static sonar images. Combined with a few other techniques, such as a video event classifier, weighted loss, and confidence threshold, it results in a more accurate and generalised system that can be used to solve a real-world problem. Then, we explore the idea of expanding the training set directly in feature space. We apply virtual adversarial training to video descriptors and observe an improvement in performance in the unsupervised video domain adaptation setting. Finally, we pay close attention to the temporal domain and utilise it to introduce a large number of video-specific transformations, along with a magnitude augmentation framework. The obtained results show that time domain consideration while designing video transformations is beneficial for networks that aim to solve video action recognition.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Jennifer Whitaker
Date Deposited:	20 Feb 2024 13:50
Last Modified:	20 Feb 2024 13:50
URI:	https://ueaeprints.uea.ac.uk/id/eprint/94365
DOI:

Downloads

Actions (login required)

View Item