Unsupervised Intra-domain Adaptation for Semantic Segmentation

Unsupervised Intra-domain Adaptation for Semantic Segmentation 을 읽고 간략하게 정리한 글입니다.

---

Unsupervised Intra-domain Adaptation for Semantic Segmentation

• automatically annotated data has a problem.
  • synthetic data -> real data
  • directly adapting models from the source data to the unlabeled target data (to reduce the inter-domain gap)
  • But result? ==> bad :(
    • there is the large distribution gap among the target data itself(intra-domain gap)

Approach

two-step self-supervised domain adaptation approach to minimize the inter-domain and intra-domain gap together.

1. inter-domain gap

   separate the target domain into an easy & hard split (using entropy-based ranking function)

2. intra-domain gap

   self-supervised adaption from the easy to hard split

   • segmentation predictions of easy split data (from G_inter) => pseudo labels 로 사용
   • Given easy split data & pseudo labels, hard split data => D_intra는 easy? hard? 판별

Results