an adversarial learning scheme for semi-supervised semantic segmentation
    sparkle： 提出对抗+SSL
    Limitations ：
    （1）require pre-defined categories, which is a general property of classification-based tasks
    （2）D averages all predicted pixels to get a single confidence value during training, as its target is an image-level real or fake probability. Using an average confidence to represent the overall confidence is not appropriate. GCT用了flaw detector来解决GCT
    （3） harder to train.
    image.png
    ~P8]1Y`WS$A{JR_LZAD~3IA.jpg

    Discriminator network：
    image.png
    Ladv(segmentation network) ：learn higher order structural information without post-processing.
    Segmentation network：
    image.png
    对于unlabel，加了Lsemi
    image.png
    Tips：
    Ladv既能fool discriminator让S输出与y相似，也让labeled学到了更深层的特征