Saliency-guided mixup with optimal re-arrangements for efficient data augmentation

Abstract

The present disclosure provides methods, apparatuses, and computer-readable mediums for performing data augmentation. In some embodiments, a method of performing data augmentation by a device includes obtaining a plurality of images from a dataset. The method further includes computing, for each image of the plurality of images, a corresponding saliency map based on a gradient of a full loss function of that image. The method further includes selecting, from a subset of arrangements of a plurality of possible arrangements, a rearrangement offset that maximizes an overall saliency of a resulting image combining the plurality of images. The method further includes generating, using the rearrangement offset and a plurality of mixing ratios, a new mixed image from the plurality of images and a new mixed label from corresponding labels of the plurality of images. The method further includes augmenting the dataset with the new mixed image and the new mixed label.