Cluster-based Absent Label Completion for Image-Text Retrieval
Xiaohan JIANG , Hailang HUANG , Zhijie NIE , Junfan CHEN , Richong ZHANG
Recent image–text retrieval approaches have achieved impressive performance, with most methods leveraging annotated image–text pairs and training models via supervised contrastive learning. However, existing methods suffer from two notable limitations: the missing positive-label problem, i.e., relying on a limited number of annotated pairs fails to provide sufficiently strong supervision; and the incorrect negative-label problem, i.e., unannotated image–text pairs are treated as mismatches during training even though some of them may actually match. To address these issues, we propose Cluster-based Absent Label Completion, a novel model that incorporates semantic structure discovered by clustering into the learned embedding space to complete missing labels. The proposed model introduces two key losses: Positive Label Completion and Negative Label Completion. Positive Label Completion identifies incomplete positive image–text pairs and completes them based on cluster assignments, thereby improving retrieval performance. Meanwhile, Negative Label Completion uses the same cluster structure to complete missing negative labels and further mitigates false negatives by leveraging supervised signals. In addition, our model is designed as a plug-and-play module and can be readily integrated into existing image-text retrieval models. Extensive experiments across multiple image-text retrieval models and datasets demonstrate that model consistently boosts the model performances and achieves state-of-the-art results.
image-text retrieval / absent label completion / clustering
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(8705KB)
