Adversarial Learning for Semi-Supervised Semantic Segmentation(BMVC2018)

Adversarial Learning for Semi-Supervised Semantic Segmentation

本文核心要点如下

• SegNet as Generator
• Discriminator: use a FCN(HxWxC -> HxWx1)
• Semi-supervised: 对于unlabeled data,
  
  • adv loss, (fix D, maximize D(S(X)))
  • self-taught, 使用D的结果作为psedo mask,去训练SegNet

D

FCN, output the probability of mask for each location
train use labeled data
sum over all position

G(S)

labeled data: cross-entropy loss(with gt), adv loss
unlaeled data: semi loss(with psedo gt), adv loss

hyperpara的没有更深入的分析