Understanding scene structure from images

The task of recovering 3D information from 2D images has long been a focus of Computer Vision research. Such information is useful in many applications: from robot navigation, where it allows the robot to understand the physical constraints of the environment it is in, to augmented reality, where 3D information is used to alter images and videos in physically plausible ways. While much progress has been made in this line of research there is still scope for further improvement. This is especially true in the case of pictures taken "in the wild", where variables such as the presence of clutter, people, irregularly shaped buildings, unusual camera angles, etc tend to cause current techniques to fail. In this work we focus on recovering 3D information from images in the presence of clutter and other such variables. We work on both indoor and outdoor scenes, utilizing different approaches in each case in order to make the 3D information recovery more robust. Since this work focuses on expanding existing techniques to work well on more challenging datasets, we had to create new datasets for both indoor and outdoor scenes that could test the robustness of our methods. Details of these datasets are also provided in this work