Feature Alignment and Contrastive Learning for Dense Instance Segmentation

Instance segmentation is an important but challenging computer vision task. Existing approaches struggle to recognize independent objects in dense scenarios where objects are tightly in contact. In this paper, we propose a segmentation network called FAC-Mask. On the one hand, it incorporates a feature alignment module, which facilitates the network to effectively integrate multi-scale information. On the other hand, we design a contrastive learning module to assist the segmentation head in feature representation learning, by learning common features among similar instances and distinguishing differences between non-similar instances. We conduct experiments on the Cityspace dataset, our model improves the detection and segmentation accuracy by +2.4% and +2.8% when fused with RefineMask. The results have demonstrated the efficacy of our method.