Transformers have achieved great success in the field of computer vision, such as PP-MobileSeg, a semantic segmentation network on mobile devices, which captures rich semantic information and details through its multi-stage StrideFormer backbone. Despite PP-MobileSeg has shown great performance, its deep network branch captures the global information insufficiently. The fusion module is relatively simple, simply entering the segmentation head for convolution and other operations and it cannot fully fuse multi-scale features. Global information can assist in better describing local semantic information, which is crucial for segmentation. GCBlock integrates global context information into the channel, effectively extracting global contextual information and is lightweight. In this paper, we propose a Fast and Accurate Semantic Segmentation Network on Mobile Devices with Global Context Modeling (GC-MobileSeg). Specifically, we introduce global context modeling into deep network branch of PP-MobileSeg, which can effectively extract global context information, making semantic segmentation accurate and fully extracting global information. Extensive experiments on Cityscapes dataset and ADE20k dataset demonstrate the effectiveness of our proposed GC-MobileSeg. We have obtained 40.20%% mIoU and 2.16(G)FLOPs on ADE20K dataset with base setting, with negligible increase in the number of parameters. 77.58% mIoU and 4.05(G)FLOPs were also obtained on the Cityscapes dataset and base setting, with negligible increase in parameter count.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    GC-MobileSeg: Fast and Accurate Semantic Segmentation Network on Mobile Devices with Global Context Modeling


    Contributors:


    Publication date :

    2023-10-11


    Size :

    3270997 byte




    Type of media :

    Conference paper


    Type of material :

    Electronic Resource


    Language :

    English