Using Co-Occurrence and Segmentation to Learn Feature-Based Object Models from Video