"The Role of Shape in Object Recognition"

Discriminative tasks, including object categorization and detection, are central components of high-level computer vision. Shape is one component of object classes that is often used to recognize or localize instances from the class in question. Many existing methods use implicit shape to guide the search for instances from various object classes. In this talk, I will discuss the role of shape in high-level computer vision, and in particular will focus on the case where we are interested in more refined aspects of the object in an image, such as pose or articulation. In these cases, a more explicit representation of the shape is preferred. In this talk I will present a method (LOOPS) for learning a shape and image feature model that can be trained on a particular image class, and used to outline instances of the class in novel images. While the training data consists of uncorresponded outlines, the resulting LOOPS model contains semantically consistent landmark points that can be localized in an image. This localization facilitates a number of tasks beyond localization, including classification along the shape axis, in which a very small number of training instances are labeled. For example, we might distinguish between cheetahs that are running and those standing still. From this small number of labeled instances, we can use our localized outlines together with a simple nearest neighbor classifier to label novel test images with the label of interest.