Towards Domain Generalisation in Artificial Intelligence

Despite the outstanding performance of modern deep learning models in various day-to-day applications, these deep learning models are typically designed under the i.i.d. assumption, where they are trained and evaluated on data sampled from the same (source) distribution. However, in real-world deployment, target distributions often differ from source data, leading to substantial performance degradation. This issue is known as domain shift. Domain Generalisation (DG) seeks to bridge the gap induced by domain shift through enabling models to generalise to Out-of-Distribution (OOD) data without accessing the target distributions during training, hence, enhancing their robustness to unseen conditions. In this thesis, we examine the problem within the context of computer vision, with particular emphasis on the tasks of image classification and object detection in real-life applications such as remote sensing and earth observations using satellite imagery. The thesis aims to address existing gaps in the literature related to the analysis and understanding of the problem and the development of effective solutions. Furthermore, despite the need, standardised DG benchmark datasets specifically designed for assessing these models under spatial domain shift (i.e., covariate shift) at a global scale is currently lacking. Therefore, in this thesis, we propose Domain Shift across Geographic Regions (DSGR) and Real-World Distribution Shifts (RWDS) datasets and benchmarks for land-use classification and various applications in object detection respectively. With these datasets and benchmarks, our objective is to provide a comprehensive analysis of the OOD performance of current deep learning techniques applied to the aforementioned tasks. We aim to shed light on their limitations and outline future research directions toward the development of more advanced DG methods. Moreover, this work seeks to enable effective evaluation of the generalisability and robustness of deep learning models in real-world applications.