Scholarlink

<p>Recent advances in deep supervised hashing have made remarkable achievements in the large-scale image retrieval task. However, the main training paradigms (<i>i.e.</i>, pairwise and pointwise) of existing deep supervised hashing methods, which will significantly impact the performance of deep supervised hashing methods under practical retrieval tasks, remain insufficiently explored. Motivated by the critical role of training paradigms in deep supervised hashing and the lack of comprehensive evaluations in this area, we systematically establish the evaluation protocols and conduct an extensive study through 1,833 experiments, yielding 7,332 results across 12 datasets. Our key findings include observations such as: 1) Pointwise hashing methods tend to exhibit higher retrieval accuracy in scenarios with seen-class queries but underperform significantly with unseen-class queries. 2) Pointwise hashing methods show greater robustness with seen-class queries, whereas pairwise hashing methods with soft constraints excel when queries are from unseen classes. 3) The impact of hash code dimensions is minimal on the retrieval performance of pointwise hashing methods but more pronounced for pairwise hashing, primarily due to suboptimal real-valued feature optimization. Code along with training logs for all experiments are open-source and available at <a href="https://github.com/aassxun/DSH_Analysis">https://github.com/aassxun/DSH_Analysis</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Empirical Study on Training Paradigms for Deep Supervised Hashing

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Yang Shen,
  • Peng Wang,
  • Xiu-Shen Wei,
  • Yazhou Yao

摘要

Recent advances in deep supervised hashing have made remarkable achievements in the large-scale image retrieval task. However, the main training paradigms (i.e., pairwise and pointwise) of existing deep supervised hashing methods, which will significantly impact the performance of deep supervised hashing methods under practical retrieval tasks, remain insufficiently explored. Motivated by the critical role of training paradigms in deep supervised hashing and the lack of comprehensive evaluations in this area, we systematically establish the evaluation protocols and conduct an extensive study through 1,833 experiments, yielding 7,332 results across 12 datasets. Our key findings include observations such as: 1) Pointwise hashing methods tend to exhibit higher retrieval accuracy in scenarios with seen-class queries but underperform significantly with unseen-class queries. 2) Pointwise hashing methods show greater robustness with seen-class queries, whereas pairwise hashing methods with soft constraints excel when queries are from unseen classes. 3) The impact of hash code dimensions is minimal on the retrieval performance of pointwise hashing methods but more pronounced for pairwise hashing, primarily due to suboptimal real-valued feature optimization. Code along with training logs for all experiments are open-source and available at https://github.com/aassxun/DSH_Analysis.