Piecing and Chipping: An effective solution for the information-erasing view generation in Self-supervised Learning
In self-supervised learning frameworks, deep networks are optimized to align different views of an instance that contains the similar visual semantic information. The views are generated by conducting series of data augmentation to the anchor samples. Although the data augmentation operations are often designed to be aggressive and extensive to lower the mutual information between views, the family of Information-Erasing data augmentation that masks out region of images is barely considered. In this work, we propose the Piecing and Chipping enhanced Erasing Augmentation (PCEA) approach to making the self-supervised learning algorithms benefit from the effectiveness of Information-Erasing data augmentation. Specifically, we design a pipeline to generate mutually weakly related transformed views using random erasing and build corresponding loss terms to take advantage of these views. Extensive experiments demonstrate the effectiveness of our method. Particularly, applying our PCEA to MoCo v2 improves the baseline by 12.84\%, 3.3\% in terms of linear classification on ImageNet-100 and ImageNet-1K.
PDF Abstract
ImageNet
MS COCO
