Entropy-weighted reconstruction adversary and curriculum pseudo labeling for domain adaptation in semantic segmentation

摘要

Supervised learning has dominated and gradually evolved into a de facto choice for automated semantic segmentation tasks, in spite of the need for large corpora of pixel-level annotated real-world data. Domain adaptation (DA), especially methods based on adversarial machine learning, can lessen labeling costs by securing that knowledge learned from synthetic (source) data with free annotations is transfer-able to unlabeled real-world data. However, commonly employed approaches in adversarial-based DA merely consider the correlation alignment between the source and target domains while completely neglecting the uncertainty information introduced from the generator. In this paper, we aim to alleviate this issue by introducing a novel entropy-guided adversarial learning framework with the reconstruction error constraint. Firstly, a perceptual-based color space is employed to transform the synthetic source images into the desired new space to approximate the appearances of the real-world images in the target domain. Secondly, an entropy-weighted adversarial framework is designed to enhance the discriminativeness and transferability of the presented model to the target domain via an autoencoder-based discriminator. In addition, a dynamic pseudo-labeling mechanism is introduced to work in conjunction with the curriculum-based self-training strategy to further improve the domain adaptability of the model. Experimental results on two well-known DA benchmarks demonstrate that the proposed method outperforms existing similar approaches in the task of semantic segmentation. (C) 2022 Elsevier B.V. All rights reserved.