Gifts from Gallery: Advancing Image Retrieval via Unsupervised Asymmetric Feature Fusion
Wengang Zhou, Hui Wu, Min Wang, Houqiang Li · IEEE Transactions on Multimedia · 2025
Asymmetric retrieval systems, characterized by the deployment of models with varying capacities on platforms with differing computational and storage resources, pose a challenge in balancing retrieval efficiency and accuracy. The recent introduction of the Asymmetric Feature Fusion (AFF) paradigm has shown promise by enhancing existing asymmetric retrieval systems through feature fusion on the gallery side. However, its reliance on extensive human-annotated data hinders practical applicability. To this end, we introduce an innovative unsupervised training method tailored for AFF. Leveraging multiple gallery models as feature extractors, our approach exploits similarities among images encoded by these models as pseudolabels. For each gallery model, we calculate its adaptive weight through an in-depth exploration of contextual relationships among ranking list images in its embedding spaces. Subsequently, these weights are utilized to effectively fuse the image similarities encoded by different gallery models into a more robust one. The fused image similarities provide powerful pseudo-supervision for training AFF. Our unsupervised training approach enhances the generality and utility of AFF in real-world scenarios, particularly when labeled data is limited or expensive to obtain. Exhaustive experiments on various landmark retrieval datasets demonstrate the superiority of our method.