Semantic segmentation with efficient tree-based methods

Giving a meaning to an image is often only possible if object positions or even the object boundaries are available. In remote sensing there often is the task to divide an aerial or satellite image into previously defined parts like forest, grassland, streets and city. Furthermore, for autonomously driving cars it might be interesting to analyze the surrounding. Methods solving the problem of automatically labeling images pixel-wise are called Semantic Segmentation. The following work deals with these problems and its solution. This can be done by extracting features from the local neighborhood and classifying them. In this work multiple methods for Semantic Segmentation are presented. In the first part of this work, we concentrate on improving already existing methods by replacing the classification step by methods based on Gaussian process classification. Gaussian processes are powerful tools in machine learning which offer a wide applicability in regression and classification problems due to their non-parametric and non-linear behavior. However, one of their main drawbacks is the training complexity which scales cubically with the number of samples. Due to the high amount of samples in an image, we present two ideas to bypass this problem. Furthermore, we present two ways to extend GrabCut, an unsupervised segmentation method, to Semantic Segmentation. These two methods are specialized to extract a single class object and its background in an image. The benefit of those methods is that incorporating shape information is very simple. The main contribution of this work focuses on Iterative Context Forests, which are based on random decision forests. This new method is fast and the computation time can be adjusted with respect to the needs of the user. It is very flexible and can be used for several tasks like classification of street scenes, images, or satellite images. However, the usage of contextual cues is the main advantage of Iterative Context Forests. It models not only local, but also a larger neighborhood up to the whole image. Therefore, misclassifications like a blue car, being classified as sky, can be prevented. Previous methods with this qualification, like Markov Random Field based approaches, are very slow in comparison to ours. Finally, extensive experiments on multiple datasets are done and show the advantages of our proposed methods.