Whether in medicine for tomography, face recognition, or video surveillance, image segmentation plays a significant role in various disciplines. Image segmentation is a crucial ongoing problem in computer vision. It aims to classify the image into parts to better analyze and understand it.
Recently, deep learning models achieved excellent results, but several other models are also achieving good results, e.g., SVMs, Auto-Encoder, or Gaussian processes. The major advantage of using Gaussian processes is that we can estimate the uncertainty of a regression or a classification. That opens many new possibilities (e.g. more precise boosting). This thesis discusses the benefits and drawbacks of a treed Gaussian process, the theoretical properties of decision trees and Gaussian processes, and how treed Gaussian processes combine these two concepts. Then we evaluate the performance on a customized MNIST data set.