UC San Diego

This thesis presents a set of novel algorithms that address practical limitations in existing object and scene understanding systems. These limitations include high computational demands of the systems, the lack of unified models that can model multiple tasks accurately in an efficient manner, and the high dependency on output labels which are in many cases difficult and expensive to collect. Our strategy is as follows: We first identify important and intuitive structures in the respective problems. Then, we analyze the existing architectures and introduce additional dimensions of variations in the computational and model structures. The proposed algorithms are more "adaptive'' than their respective baselines, in the sense that they can now use the new degrees of freedoms to address the limitations in the latter. These algorithms are practical because they demonstrate significant empirical successes in addressing the limitations of existing methods. They have pushed the state-of-the-art and inspired follow-up studies in designing better object and scene understanding systems for real-world challenges.