Unsupervised Intra-Domain Adaptation
CNN-based neural networks heavily rely on annotated data which are labor intensive.
To cope with this limitation, automatically generated data are used to train segmentation models. However, models trained with these data struggle to generalize to real images.
To tackle this challenge, many researchers have tried to minimize the gap between the training data and the target data(called the “inter-domain gap” in this paper), but they have failed to adapt to the gap within the target data(called the “intra-domain gap”).
This paper proposes a solution for this.
Architecture
There are mainly 3 stages.
- Inter-Domain Adaptation
Uses the common unsupervised domain adaptation method. The GENERATOR will generate an image to fool the DISCRIMINATOR, and the DISCRIMINATOR will label the domain name of the generated image. - Entropy-based ranking
By ranking the images from the GENERATOR, the model can divide the data within the target domain into “easy” and “hard” samples. - Intra-Doman Adaption
Trained in the same way as the Inter-Domain Adaption phase, using “hard” samples as its new target domain.
Reference:Unsupervised Intra-Domain Adaptation for Semantic Segmentation Through Self-Supervision
