Few-shot semantic segmentation: a review on recent approaches

Zhaobin Chang, Yonggang Lu, Xingcheng Ran, Xiong Gao, Xiangwen Wang

Abstract: Few-shot semantic segmentation (FSS) is a challenging task that aims to learn to segment novel categories with only a few labeled images, and it has a wide range of real-world applications. Recently, the performance of FSS has been greatly promoted by using deep learning approaches. In this paper, we provide a systematic review of recent advances to fully understand FSS. Firstly, we introduce the definition and evaluation metrics of FSS as well as popular datasets. Next, we review representative FSS approaches and categorize them into FSS based on parametric metric learning and FSS based on non-parametric metric learning from the parameter learning perspectives in the metric phase. With this taxonomy, we summarize and discuss the pros and cons of different FSS methods, and quantitative results are given for the described methods, following up with a discussion of the results. We then introduce some representative applications of FSS. Finally, we discuss the limitations of the existing approaches and provide potential future research directions for FSS.