Double Similarity Distillation for Semantic Image Segmentation

The balance between high accuracy and speed has always been a challenging task in semantic image segmentation. Compact segmentation networks are more widely used the case of limited resources, while their performances constrained. In this paper, motivated by residual learning global aggregation, we propose simple yet general effective knowledge distillation framework called double similarity (DSD) to improve classification all existing compact capturing pixel category dimensions, respectively. Specifically, pixel-wise (PSD) module that utilizes attention maps capture detailed spatial dependencies across multiple layers. Compared with exiting methods, PSD greatly reduces amount calculation is easy expand. Furthermore, considering differences characteristics other computer vision tasks, category-wise (CSD) module, which can help network strengthen correlation constructing matrix. Combining these two modules, DSD no extra parameters only minimal increase FLOPs. Extensive experiments on four datasets, including Cityscapes, CamVid, ADE20K, Pascal VOC 2012, show outperforms current state-of-the-art proving its effectiveness generality. code models will be publicly available.