介绍骨架
    The subsequent developments studies the models from three main aspects: resolution, context, and edge.
    The works on resolution enlargement include mediating the spatial loss caused in the classification network, e.g., using the encoder-decoder scheme or dilated convolutions, and maintaining high resolution, such as HRNet.
    The works on exploiting contexts include spatial context, e.g. PSPNet and ASPP, object context and application of self-attention.
    Improving the segmentation quality on the edge areas include Gated-SCNN, PointRend, and SegFix.