In this work, we first observe that for semantic segmentation, the low density regions are more apparent within the hidden representations than within the input.

sparkle

（1）hidden representations
（2）比较了various types of perturbations

image.png
image.png
image.png
image.png
image.png

Perturbation functions

three types of perturbation functions：feature based, prediction based and random.

Feature based perturbations

injecting noise into or dropping some of the activations of encoder’s output feature map z.
F-Noise ：给z加高斯噪声
F-Drop ：随机丢掉一些z

Prediction based perturbations

adding perturbations based on the main decoder’s prediction yˆ = g(z) or that of the auxiliary decoders
Guided Masking ： 网络可能过于依赖上下文关系，遮了MOBJ让y运用 context mask Mcon = 1 − Mobj
Guided Cutout： 每个object的边界框内的随机裁剪
Intermediate VAT ： 虚拟对抗训练image.png

Random perturbations

DropOut