Distilling Localization for Self-Supervised Representation Learning

Recent progress in contrastive learning has revolutionized unsupervised representation learning. Concretely, multiple views (augmentations) from the same image are encouraged to map close embeddings, while different images pulled apart.In this paper, through visualizing and diagnosing classification errors, we observe that current models ineffective at localizing foreground object, limiting their ability extract discriminative high-level features. This is due fact view generation process considers pixels an uniformly.To address problem, propose a data-driven approach for invariance backgrounds. It first estimates saliency then creates augmentations by copy-and-pasting onto variety of back-grounds. The still follows instance discrimination approach, so trained disregard background content focus on foreground. We study estimation methods, find most methods lead improvements With significant performance achieved self-supervised ImageNet classification, also object detection PASCAL VOC MSCOCO.