Going into depth Learning morphological aspects in data modalities using neural networks

In this thesis, we recognise that different data modalities have specific algebras describing them.
Specifically, we are interested in non-linear geometric 'contact' signals of which depth maps are the prime example. This thesis has sought to explore a variety of techniques for the analysis of depth modalities by means of neural networks.
The first part of the thesis addresses predicting depth from single images using image-to-image reconstruction. This problem is inherently ill-defined, and requires the use of many loss terms to converge to a reasonable solution. The usual method of weighting losses manually is a time-consuming and computationally expensive process, and can be much sped up using the proposed statistics-based automated loss weighting method.
The second part of the thesis investigates the merit of changing the core building blocks of neural networks to mimic the underlying algebra of the depth data. Training and using morphological networks is not straightforward, since foundational algorithms such as back-propagation have to be revisited and low-level operators have to be implemented. In the process, we found that many standard constituents of linear networks, such as pooling and ReLU, are morphological in nature. We conclude that the performance of morphological networks compared to their linear counterparts generally increases, and moreover at reduced parameter counts in a variety of depth-related tasks.