DECNet: Dense embedding contrast for unsupervised semantic segmentation

Xiaoqin Zhang, Baiyu Chen, Xiaolong Zhou, Sixian Chan

Abstract: Highlights•A simple contrastive framework is advanced for unsupervised semantic segmentation.•The NNS strategy is proposed to build well-defined positives and negatives.•A new objective is designed to mitigate the inherent defect of contrastive learning.•Extensive experiments show that proposed approach outperforms state-of-the-arts.