Urban area road scene segmentation using InternImage-adapter model

Abstract

With the increasing development of autonomous driving, there is a need for an accurate deep learning model to detect and segment important objects such as other vehicles, traffic lights, road signs, road segments, pedestrians, and drivers accurately. Urban road scene segmentation presents the greatest challenge due to the presence of numerous obstacles, including pedestrians, roadside vegetation, buildings, and various other elements. The proposed model is a combination InternImage as the backbone and taking the Adapter of ViT-Adapter applied in the first block of the backbone model. The outputs of the output of InternImage last block and adapter output will be combined by element-wise addition then fed to segmentor. The model evaluated by public dataset Cityscapes with mIoU as the measuring metric. The result achieved is 81.93 mIoU on test data. The addition of adapter on the first block to InternImage does improve the performance of standalone InternImage.