A Semantic Segmentation Method Based on Multilevel Relational Distillation

This article proposes a multilevel relational distillation method, which focuses more on the spatial structure relationship of feature maps compared to traditional detection distillation. Compared with advanced network models, it has a more obvious segmentation effect on small targets in complex scenes. On the ADE20K dataset, the accuracy reached 75.93%, and on the Cityscapes validation set, the average intersection to union ratio was 74.12%, proving that the model has improved on various datasets and has strong universality.