Structured serialization semantic transfer network for unsupervised cross-domain and retrieval

Unsupervised domain adaptation aims to apply a model trained from labeled datasets to unlabeled datasets, which still exist challenges to minimize the degree of inter-domain sam-ple differences. In the paper, we design the structured serialization semantic transfer net-work (S3TN) to further minimize the domain gap by structured and serialization information at the semantic level. Our S3TN contains the spatially structured alignment module (SSA) and the category serialization alignment module (CSA). The spatially structured alignment module preserves the structured information between different category-level centers (spatial associations between different categories in the feature space of category-level centers), further aligning the structured correlations between different domains. The serialization alignment module enhances the reliability of the category-level centers by avoiding valid information loss and error information retention, which further mitigates the problem of knowledge forgetting. Various experiments have been conducted on three widely-used datasets, Office-31, MI3DOR-1, and MI3DOR-2, and the experimental results verify the effectiveness of our proposed method. Especially on Office-31 dataset, our method outperforms the existing methods with gains of 1.3% for the average of the classification accuracy.

Automated summary

Unsupervised domain adaptation aims to minimize inter-domain sample differences. This paper proposes a structured serialization semantic transfer network (S3TN) to further reduce the domain gap by utilizing structured and serialization information at the semantic level. Experimental results show that the proposed method achieves significant improvements in classification accuracy.