1: \begin{abstract}

2:   We present {\tt ReCo}, a contrastive learning framework designed at a regional level to assist learning in semantic segmentation. ReCo performs semi-supervised or supervised pixel-level contrastive learning on a sparse set of hard negative pixels, with minimal additional memory footprint.

3:   ReCo is easy to implement, being built on top of off-the-shelf segmentation networks, and consistently improves performance in both semi-supervised and supervised semantic segmentation methods, achieving smoother segmentation boundaries and faster convergence.  The strongest effect is in semi-supervised learning with very few labels. With ReCo, we achieve high quality semantic segmentation models, requiring only 5 examples of each semantic class. Code is available at \url{https://github.com/lorenmt/reco}.

4: \end{abstract}

5: