Coarse-to-fine label refinement for domain adaptive retrieval

Domain adaptive retrieval (DAR) is a promising research field. However, existing methods still suffer from the following limitations: 1) they rely heavily on pseudo-labeling strategies that oversimplify complex relationships between samples; 2) they treat labels as algorithmic outputs rather than optimizable variables, potentially breaking some natural connections between features and categories. To address these issues, we propose an effective approach called Coarse-to-Fine Label Refinement (CFLR). First, joint orthogonal matrix factorization is employed: one is to learn an optimizable latent feature representation, the other is to decompose predefined coarse pseudo-labels into improvable continuous values. Second, a classifier is introduced to connect these components, establishing a mutually reinforcing relationship between features and labels. This mutual enhancement captures implicit cross-category semantics by mining the iteratively updated feature information. Based on the refined labels, we develop an improved graph embedding that achieves more natural cross-domain relationships. Finally, high-quality hash codes are generated by directly quantifying the refined semantics. Experiments on multiple popular cross-domain benchmark datasets demonstrate that the proposed CFLR achieves state-of-the-art performance.