ESL: Entropy-guided Self-supervised Learning for Domain Adaptation in Semantic Segmentation

While fully-supervised deep learning yields good models for urban scene\nsemantic segmentation, these models struggle to generalize to new environments\nwith different lighting or weather conditions for instance. In addition,\nproducing the extensive pixel-level annotations that the task requires comes at\na great cost. Unsupervised domain adaptation (UDA) is one approach that tries\nto address these issues in order to make such systems more scalable. In\nparticular, self-supervised learning (SSL) has recently become an effective\nstrategy for UDA in semantic segmentation. At the core of such methods lies\n`pseudo-labeling', that is, the practice of assigning high-confident class\npredictions as pseudo-labels, subsequently used as true labels, for target\ndata. To collect pseudo-labels, previous works often rely on the highest\nsoftmax score, which we here argue as an unfavorable confidence measurement.\n In this work, we propose Entropy-guided Self-supervised Learning (ESL),\nleveraging entropy as the confidence indicator for producing more accurate\npseudo-labels. On different UDA benchmarks, ESL consistently outperforms strong\nSSL baselines and achieves state-of-the-art results.\n