Investigation and Improvements of Neural Field Efficiency and Quality

Surface reconstruction is a technique in the field of computer graphics that creates a 3D model from a set of points or other form of data. With the growing presence of machine learning and its impressive advances in different areas, representing shapes with the help of neural networks got a lot of interest under researcher. Neural fields are a specific category of neural networks that deal with coordinate-based input. They use a neural network as a compact representation for objects represented as continuous functions of spatial coordinates, such as an image or 3D shape. This thesis is concerned with the investigation of existing neural field models for their performance and limitations in the context of 3D shape representation. Analyzing their strength and weakness, for this instance, and develop modifications to existing methods and models. From there, create an alternative neural field model and improve the process for solving such a task. To achieve this, this work presents the additive progressive detail/displacement network. A network architecture that combines techniques from different neural field concepts and implements a coarse-to-fine hierarchy that starts from a low detail base shape and progressively learns the displacement "stage-wise", by increasing the frequency and capacity (resolution) throughout the networks. Such an approach is highly customizable and tries to better capture the 3D surface of more complex shapes or structures by splitting the workload to multiple networks. This also allows updating and revisiting individual stages of the model to enhance the detail capture of an object surface in regions of interest. Furthermore introducing traditional geometry-driven methods when working with oriented point clouds into the machine learning domain, like adaptive sampling strategies that better distribute the necessary training data across the networks. This involves a type of importance sampling that provides data not just when, but also where it matters in the network hierarchy and pipeline. Focusing on cutting down unnecessary samples, lowering computational expenses and helping the network better understand the data, which can lead to improved results and is also memory efficient. The proposed system in this work is tailored to reconstructing 3D surfaces, but is not limited to other signal encodings in lower or higher dimensional space due to the nature of neural networks as an universal function approximator. With regards to a potential application, like additive manufacturing, the work also provides a user interface for quick experiments and previews when training and testing the model on different kinds of shapes. This allows rapid prototyping further increasing the potential usage of neural fields in a more industry oriented workflow.