Coherent Occlusion Reasoning for Instance Recognition

Occlusions are common in real world scenes and are a major obstacle to robust object detection. In this paper, we present a method to coherently reason about occlusions on many types of detectors. Previous approaches primarily enforced local coherency or learned the occlusion structure from data. However, local coherency ignores the occlusion structure in real world scenes and learning from data requires tediously labeling many examples of occlusions for every view of every object. Other approaches require binary classifications of matching scores. We address these limitations by formulating occlusion reasoning as an efficient search over occluding blocks which best explain a probabilistic matching pattern. Our method demonstrates significant improvement in estimating the mask of the occluding region and improves object instance detection on a challenging dataset of objects under severe occlusions.