Hierarchical feature disentangling network for universal domain adaptation

Highlights：

• We propose a novel Hierarchical Feature Disentangling Network (HFDN) for universal domain adaptation (UniDA), which is a more practical domain adaptation (DA) setting compared with close-set, open-set, and partial DA.

• The proposed HFDN is the first to address the feature misalignment problem caused by both the domain gap and the category gap in UniDA.

• With the disentangled features, we propose an innovated knowledge transfer approach for UniDA, which bridges the domain gap to reduce the domain-shift by domain adversarial training, and leverages information of category gap to assign larger weights for samples from the common label set.

• Experiments demonstrate that our method achieves the best performance for UniDA on benchmark datasets compared with existing methods proposed under various domain adaptation assumptions.