Before their second birthday, human toddlers master the dynamics of their surroundings better than any machine. For example, when an object goes out-of-sight, such as a due to an occlusion, children understand the object to still exist. However, despite rapid progress in deep learning, these foundations of visual intelligence are missing from computer vision, undermining key applications in situational awareness across health, security, and robotics. The goal of this summer project is to create computer vision systems that track visual objects through occlusions. The student will investigate machine learning algorithms that learn object permanence without human supervision by leveraging both unlabeled video and interaction with its environment.
