Learning to Predict Visibility and Invisibility from Occlusion Events