Geometric Supervision and Deep Structured Models for Image Segmentation

Abstract: The task of semantic segmentation aims at understanding an image at a pixel level. Due to its applicability in many areas, such as autonomous vehicles, robotics and medical surgery assistance, semantic segmentation has become an essential task in image analysis. During the last few years a lot of progress have been made for image segmentation algorithms, mainly due to the introduction of deep learning methods, in particular the use of Convolutional Neural Networks (CNNs). CNNs are powerful for modeling complex connections between input and output data but have two drawbacks when it comes to semantic segmentation. Firstly, CNNs lack the ability to directly model dependent output structures, for instance, explicitly enforcing properties such as label smoothness and coherence. This drawback motivates the use of Conditional Random Fields (CRFs), applied as a post-processing step in semantic segmentation. Secondly, training CNNs requires large amounts of annotated data. For segmentation this amounts to dense, pixel-level, annotations that are very time-consuming to acquire. This thesis summarizes the content of five papers addressing the two aforementioned drawbacks of CNNs. The first two papers present methods on how geometric 3D models can be used to improve segmentation models. The 3D models can be created with little human labour and can be used as a supervisory signal to improve the robustness of semantic segmentation and long-term visual localization methods. The last three papers focuses on models combining CNNs and CRFs for semantic segmentation. The models consist of a CNN capable of learning complex image features coupled with a CRF capable of learning dependencies between output variables. Emphasis has been on creating models that are possible to train end-to-end, giving the CNN and the CRF a chance to learn how to interact and exploit complementary information to achieve better performance.