Data Augmentation With Flickering Backgrounds and Instances for Instance Segmentation

Kisu Lee, Jungeun Kim, Ha Young Kim

Abstract: Data augmentation (DA) tailored to instances is vital for instance segmentation to improve model robustness and accuracy without high manual annotation costs. Existing erasing methods risk losing information about instances, whereas instance-level methods necessitate additional overhead, such as an object bank and context calculations to decide locations to attach objects to an image. Thus, we propose the instance-centric erasing-based DA method, FlickBI, which enhances focus on the target object and diversity by randomly eliminating confusing information. FlickBI consists of two separate methods: FlickBack and FlickIns. FlickBack removes the unrelated background information based on the given annotations. FlickIns stochastically deletes instances, assuming that instances within one image are independent of each other. The experiments reveal that the proposed simple yet effective method consistently enhances performance across detectors and backbones on three benchmark datasets with just a few lines of code, even if the diversity of transformed images and the number of used instances are lower. On the COCO dataset, FlickBI achieves mask mAP improvements ranging from 0.6 to 5.5. Moreover, on the Cityscapes and LVIS datasets, there is an average improvement in mask AP of +3.1 and +4.8, respectively. Furthermore, FlickBI demonstrates synergistic improvements with other instance-level DA methods.