Light Field Processing for Immersive Systems

Light Field (LF) imaging is an immersive imaging modality that has attracted increasing attention in recent decades, due to its ability of capturing both light intensity and direction information of a scene in a Four-Dimensional (4D) array, known as the 4D LF. The rich information included in 4D LFs enables the viewer to explore the scene from different perspectives, hence, enhancing depth perception and realism. However, the promise of an immersive experience comes with challenges that need to be investigated, notably in terms of processing and editing. One example of those challenges is to efficiently process/edit 4D LFs by exploiting the spatio-angular correlations while ensuring spatial accuracy and angular consistency. Therefore, this Thesis tackles this challenge through several methods that together form a pipeline to efficiently process/edit 4D LFs. At first, this Thesis proposes an efficient disparity propagation method that enables computing angularly consistent disparity maps for all LF views. Afterwards, this Thesis proposes novel over-segmentation methods that rely on disparity maps as an additional guiding feature to group corresponding pixels across LF views into spatio- angular segments. The over-segmented 4D LFs are then used as an intermediate representation that enables LF segmentation and facilitates neural style transfer propagation. In this Thesis, it has been shown that representing 4D LFs based on over-segmentation allows the usage of classical graph cut and graph neural networks to achieve efficient 4D LF segmentation. The proposed processing and editing LF methods have shown outperforming performance in several aspects, such as spatial accuracy and angular consistency.