WP Leader: ECP
We will construct models and algorithms for the recognition and localization of object categories, such as humans, tables and chairs. The challenges lie in dealing with intra-category variation in appearance and shape, in order to generalize to new examples, and also in dealing with occlusions and pose variation. We will use training data to construct the models, and efficient optimization to apply the models during testing. We will develop algorithms that recognize object categories from multiple views, while at the same time segmenting (`parsing') the surface information delivered by depth sensors into object parts. Our algorithms will be evaluated in terms of their accuracy, measured in terms of precision-recall curves that indicate the tradeoff between false positives and missed detections, and computational efficiency, measured in the frames-per-second at which the system can operate.