DARS: Data Augmentation using Refined Segmentation on Computer Vision Tasks

Data augmentation has proven to be effective in improving convolutional neural network performance. As a result, variations of data augmentation methods such as image blending, regional dropout, and copy-paste augmentations has emerged in recent years, enhancing the network's localization and generalization capabilities. However, most data augmentation approach generates unnatural image that contains visible artifacts or strong edges which the network can latch onto. In this paper, we present DARS, a module to refine segmentation in efforts to minimize artifacts for data augmentation. We demonstrated its effectiveness through object classification on Pascal VOC [6] and MS COCO [14] benchmark dataset, with qualitative comparison on refinement from DARS compared to state-of-the-art segmentation models.