RepCo: Replenish sample views with better consistency for contrastive learning

Xinyu Lei, Longjun Liu, Yi Zhang, Puhang Jia, Haonan Zhang, Nanning Zheng

Abstract: Highlights•Propose a novel SSL method that can generate high-quality image views and achieve better performance for contrastive learning.•Use similarities of sampled patches as thresholds to determine whether two views should be treated as positive or negative pairs.•The sampled positive pairs can maintain a high consistency, negative samples are semantically similar but not identical to the anchor.•Propose an indicator matrix to help capture sampled positive and negative pairs.•Provide strong benefits for downstream tasks and demonstrate competitive performance.