Title: Hashing-Baseline: Rethinking Hashing in the Age of Pretrained Models Show Chinese title

Title: 哈希基线：在预训练模型时代重新思考哈希

Abstract: Information retrieval with compact binary embeddings, also referred to as hashing, is crucial for scalable fast search applications, yet state-of-the-art hashing methods require expensive, scenario-specific training. In this work, we introduce Hashing-Baseline, a strong training-free hashing method leveraging powerful pretrained encoders that produce rich pretrained embeddings. We revisit classical, training-free hashing techniques: principal component analysis, random orthogonal projection, and threshold binarization, to produce a strong baseline for hashing. Our approach combines these techniques with frozen embeddings from state-of-the-art vision and audio encoders to yield competitive retrieval performance without any additional learning or fine-tuning. To demonstrate the generality and effectiveness of this approach, we evaluate it on standard image retrieval benchmarks as well as a newly introduced benchmark for audio hashing. Show Chinese abstract

Abstract: 使用紧凑二进制嵌入的信息检索，也称为哈希，对于可扩展的快速搜索应用至关重要，但最先进的哈希方法需要昂贵且场景特定的训练。 在本工作中，我们引入了Hashing-Baseline，这是一种强大的无需训练的哈希方法，利用产生丰富预训练嵌入的强大预训练编码器。 我们重新审视经典的、无需训练的哈希技术：主成分分析、随机正交投影和阈值二值化，以生成哈希的强大基线。 我们的方法将这些技术与最先进的视觉和音频编码器的冻结嵌入相结合，以在没有任何额外学习或微调的情况下获得具有竞争力的检索性能。 为了展示这种方法的通用性和有效性，我们在标准图像检索基准以及一个新的音频哈希基准上进行了评估。