Date of Award


Degree Type


Degree Name

Master of Science (MS)


Computer Graphics Technology

Committee Chair

Tim McGraw

Committee Member 1

David Whittinghill

Committee Member 2

Xavier Tricoche


Previous research has resulted in many representations of surfaces for rendering. However, for some approaches, an accurate representation comes at the expense of large data storage. Considering that Artifcial Neural Networks (ANNs) have been shown to achieve good performance in approximating non-linear functions in recent years, the potential to apply them to the problem of surface representation needs to be investigated. The goal in this research is to exploring how ANNs can effciently learn the Signed Distance Field (SDF) representation of shapes. Specifcally, we investigate how well different architectures of ANNs can learn 2D SDFs, 3D SDFs, and SDFs approximating a complex triangle mesh. In this research, we performed three main experiments to determine which ANN architectures and confgurations are suitable for learning SDFs by analyzing the errors in training and testing as well as rendering results. Also, three different pipelines for rendering general SDFs, grid-based SDFs, and ANN based SDFs were implemented to show the resulting images on screen. The following data are measured in this research project: the errors in training different architectures of ANNs; the errors in rendering SDFs; comparison between grid-based SDFs and ANN based SDFs. This work demonstrates the use of using ANNs to approximate the SDF to a mesh by learning the dataset through training data sampled near the mesh surface, which could be a useful technique in 3D reconstruction and rendering. We have found that the size of trained neural network is also much smaller than either the triangle mesh or grid-based SDFs, which could be useful for compression applications, and in software or hardware that has a strict requirement of memory size.