Instance Segmentation Combined CMT and Swin Transformer in Driving Scenes

Abstract

CNN and Transformer have been used widely through computer vision problems, including object detection and instance segmentation. But usually, CNN and Transformer are utilized independently. Recently, a new method called CMT has combined the advantages of both. It applies convolution to mitigate the computation overhead. In this work, we combine the advantages of CMT and swin transformer to enrich feature extraction. And build a framework that used the new backbone to achieve instance segmentation. Finally, we have done experiments in driving scenes and achieved good results.