A review on 2D instance segmentation based on deep neural networks

Image instance segmentation involves labeling pixels of images with classes and instances, which is one of the pivotal technologies in many domains, such as natural scenes understanding, intelligent driving, augmented reality and medical image analysis. With the power of deep learning, instance segmentation methods that use this technique have recently achieved remarkable progress. In this survey, we mainly discuss the representative 2D instance segmentation methods based on deep neural networks. Firstly, we summarize current fully-, weakly- and semi-supervised instance segmentation methods, and divide existing fully-supervised methods into three sub-categories depending on the number of stages. Based on our investigation, we conclude that currently, two-stage methods dominate the frontier of general instance segmentation; single-stage methods can achieve a better speed-accuracy trade-off, and multi-stage methods can achieve higher accuracy. Secondly, we introduce eleven datasets and three evaluation metrics for evaluating instance segmentation methods that can help researchers decide which one to choose to meet their needs and goals. Then the innovation and quantitative results of state-of-the-art general instance segmentation methods and specific instance segmentation methods (including salient instance segmentation, person instance segmentation, and amodal instance segmentation) are reviewed. In what follows, the common backbone networks are reviewed to better explain the reasons that why deep neural networks-based instance segmentation methods can achieve excellent performance. Finally, the future research directions and potential applications of instance segmentation are discussed, which can facilitates researchers to realize the existing technical difficulties and recent research hotspots.