核心思想:
build invariance across domains by minimizing the measure of domain shift such as correlation distances or maximum mean discrepancy
创新点(号称):
从两个角度考虑invariance: appearance-level 和 representation-level
1.appearence level (AAN)
AAN is to construct an image that captures high-level content in a source image and low-level pixel information of the target domain.
use a pre-trained CNN
2.representation level (RAN)
FCN + ASPP + domain discriminator
(ASPP的作用:Atrous Spatial Pyramid Pooling (ASPP) strategy is particularly devised to enlarge the field of view of filters in feature map and endow the domain discriminator with more power.)
RAN优化两种loss: classification loss to measure pixel-level semantics and adversarial loss to maximally fool the domain discriminator with the learnt source and target representations.
作者文中总结:The solution also leads to the elegant views of what kind of invariance should be built across domains for adaptation and how to model the domain invariance in a deep learning framework especially for the task of semantic segmentation, which are problems not yet fully understood in the literature.
In the context of domain adaptation, this adversarial principle is then equivalent to guiding the representation learning in both domains, making the difference between source and target representation distributions indistinguishable through the domain discriminator.