Augmentation Based on Artificial Occlusions for Resilient Instance Segmentation

Real-world instance segmentation applications usually demand real-time identification of objects that are small in size, occluded from other objects, appearing and disappearing in quick succession. For these reasons, instance segmentation requires a lot of representative data to grasp the subtle changes that occur in a scene. In this paper, we propose an augmentation methodology that allows for sufficient training relying on a small number of annotated data. Additionally, we provide two new datasets for instance segmentation including the semantic class firearm, for security applications. By applying the proposed augmentation technique on three datasets, the performance of instance segmentation methods is improved as indicated by the experimental results.