LPCL: Localized prominence contrastive learning for self-supervised dense visual pre-training

Zihan Chen, Hongyuan Zhu, Hao Cheng, Siya Mi, Yu Zhang, Xin Geng

Abstract: Highlights•We introduce a novel self-supervised model LPCL for dense prediction tasks.•We propose an objectness patch selection module to guide contrastive learning.•LPCL effectively improves the performance on various dense prediction tasks.