GNAH: New hashing method cuts data needs for cross-modal retrieval
Learns from just a few image-text pairs—no labels required.
A team of researchers from multiple institutions (Runhao Li, Xiaoxu Ma, et al.) introduce Global-Neighborhood Alignment Hashing (GNAH), a novel unsupervised method for cross-modal retrieval that works with limited image-text pairs. Traditional unsupervised cross-modal hashing requires large-scale unlabeled datasets, which are expensive to collect. GNAH addresses this by preserving the semantic structure of pre-trained vision-language foundation models (like CLIP) within a compact binary Hamming space.
GNAH uses two key modules: a Prototype-Anchored Global Alignment module that captures global structural information from the continuous latent space and transfers it into binary codes, and a Contrastive Stochastic Neighborhood Alignment module that extends pairwise contrastive learning by modeling stochastic neighborhood relationships. This alleviates overfitting to sparse pairwise correlations. Extensive experiments show GNAH consistently outperforms existing unsupervised cross-modal retrieval methods under data-constrained settings, offering a practical solution for real-world CMH applications where labeled data is scarce.
- GNAH requires only a small number of unlabeled image-text pairs, reducing data collection costs.
- Uses Prototype-Anchored Global Alignment to transfer global structure from latent space to binary Hamming space.
- Contrastive Stochastic Neighborhood Alignment prevents overfitting by modeling stochastic neighborhood relationships.
Why It Matters
Enables scalable cross-modal search without labeling, lowering barriers for real-world deployment.