| A key problem in computer vision is image understanding, which we define as the task of recognizing every object/region in the scene. Traditionally, this has been accomplished by considering the information within each object/region to be recognized. Incorporating contextual information, i.e. information other than the appearance information of the object, for image understanding has received significant attention in recent works. Contextual information is often learnt in a supervised manner and utilized to enhance performance of higher level tasks such as object recognition or detection. In this thesis, we take a closer look at the role of context in image understanding. Specifically, we ask three questions. First: When is context really helpful? We show, through computer vision experiments as well as human studies, that context provides improvements in recognition performances only when the appearance information is weak (such as in low resolution images or in the presence of occlusion). Second: For what tasks can contextual information be leveraged? We show that apart from high-level tasks of recognition and detection, contextual information can be effectively leveraged for low level tasks as well, such as identifying salient or representative patches in an image. Lastly, How can context be learnt? Or alternatively, how much contextual information can be extracted in an unsupervised manner? We propose a unified hierarchical representation for contextual interactions or spatial patterns among visual entities at all levels, from low-level features to parts of objects, objects, groups of objects and ultimately the entire scene. We present results of our approach on a variety of datasets such as object categories, street scenes and natural scene images. |
