Shape-centered representations

En savoir plus sur le livre

Computer vision seeks to enable machines and algorithms to 'see,' ultimately creating intelligent applications that assist humans in various scenarios. This thesis investigates computer vision across three layers: low-level features, mid-level representations, and high-level applications. Each layer builds on the previous one while imposing constraints that influence the others. The application layer integrates human-machine interfaces, linking human perception with computer vision. By examining all layers collectively, we gain deeper insights into the interplay of methods, rather than isolating problems. Constraints from different layers and users are incorporated into algorithm design, moving beyond mere algorithmic performance. Chapter 1 introduces the topic, while Chapter 2 focuses on the feature layer, presenting novel shape-centered interest points formed at locations of high local symmetry. These points demonstrate robustness against common image transformations like scaling and noise. Chapter 3 introduces two strategies for creating robust mid-level representations, including a new feature grouping method that combines the strengths of shape-centered and corner-based interest points, and a set of medial feature superpixels for compact image representation. Chapter 4 connects computer vision with human observers through three applications utilizing shape-centered representations. It presents a multi-class scene labeli