| Summary: | The need for large quantities of high quality training data is one of the overarching problems facing the Computer Vision and Deep Learning research community. The need to seek versatile, scalable solutions to this problem is imperative as neural networks become involved with almost every aspect of the modern world. The topic of this thesis is training neural networks with Synthetic Data, one of the most promising solutions to the problem of data scarcity.
In this thesis I focus these attempts on plant phenotyping tasks, an important field of interest within Computer Vision concerned with the automatic measurement of the physical features of different plants. This thesis presents a number of Synthetic Datasets created with deep learning in mind, and then details a number of novel techniques for leveraging these datasets when working on phenotyping problems, focusing on domain adaptation, style transfer and network fine-tuning.
I present a heatmap guidance extension for style transfer, and a clustering approach to deep learning training to improve generalisation on diverse target datasets. Then my work on 3D reconstruction is presented, where domain adaptation is performed simultaneously with training a volumetric regression network, allowing for an unsupervised domain adaptation approach using an unlabeled train set. I present a series of experiments comparing Synthetic Data and fine-tuning approach between CNN and Transformer based architectures. Finally I look at Diffusion Models, a new form of generative neural network that promises to be the future of synthetic data generation.
|
|---|