Addressing Long-Tail Deep Hash Retrieval via Personal Diffusion Model

In deep hash retrieval, balanced and high-quality datasets contribute to good hash functions through hash retrieval techniques, thereby generating binary hash codes that are discriminative in binary space. However, dataset of balanced distribution is usually an ideal scenario, and most realistic datasets tend to exhibit long-tailed distributions, which will greatly affect the quality of the learned hash function. This paper first observes from empirical analysis that pre-trained models in downstream tasks can greatly alleviate the impact of long-tail distribution on deep hash retrieval. Based on this finding, we propose a deep hash retrieval framework that utilizes synthetic data from the generative model for model and data enhancement respectively. In experiments, our method achieves remarkable results on both public data and complex datasets.