Vision with a Purpose: Inferring the Function of Objects in Video

Video provides the opportunity to classify vehicles, people, buildings and other structures by their behavior and the activity surrounding them. Object classes that are visually indistinguishable can be differentiated using activity, even in low-resolution video where traditional recognition from static imagery is not viable. Conversely, objects within the same class that have the same behavior but different appearance can be recognized. Our approach is based on learning probabilistic models of object function using a canonical set of low-level relations and actions computed from object tracks and scene context. We will combine traditional generative model learning, which learns models in isolation from each other, with discriminative learning that identifies how the classes are different from each other. Our recognition method will handle situations where objects are tracked for intervals shorter than their characteristic behavior duration, and longer too. In Phase 1, we will develop initial functional recognition methods for moving objects such as people and vehicles. A set of promising functional types will be identified, and we will use them to study the effects of tracking accuracy and levels of scene context on functional recognition accuracy.