Reciprocal normalization for domain adaptation

Batch normalization (BN) is widely used in modern deep neural networks, which has been shown to represent the domain-related knowledge, and thus ineffective for cross-domain tasks like unsupervised domain adaptation (UDA). Existing BN variant methods aggregate source target knowledge same channel module. However, misalignment between features of corresponding channels across domains often leads a sub-optimal transferability. In this paper, we exploit relation propose novel method, Reciprocal Normalization (RN). Specifically, RN first presents Compensation (RC) module acquire compensatory each both based on channel-wise correlation. Then develops Aggregation (RA) adaptively feature with its components. As an alternative BN, more suitable UDA problems can be easily integrated into popular methods. Experiments show that proposed outperforms existing counterparts by large margin helps state-of-the-art approaches achieve better results. The code available https://github.com/Openning07/reciprocal-normalization-for-DA.