This thesis presents a novel way of scene analysis in images and videos. Traditional scene analysis using object detection involves a lot of human labor for labeling the images, and also has the difficulty of handling a large number of objects categories. Our approach to scene analysis is unsupervised in nature. Given a video, we want to "discover" the objects of interest. No single labeled image is used to pre-train or initialize the system. Still, the system is able to discover the objects of interest. It works on a wide variety of videos and it can discover objects belonging to a large set of different categories. It works in crowded scenes with distracting background pattern and motion. It works in partial occlusions and total removal. The probabilistic framework consists of an appearance model and a motion model. The appearance model exploits the consistency of object parts in appearance across frames. The motion model exploits the motion continuity across frames. Together, they provide appearance and location estimates of the objects of interest. This framework provides a basis for higher level video content analysis tasks.
展开▼
