Hashing-Baseline: Rethinking Hashing in the Age of Pretrained Models

2025-09-17

AI Summary

Existing hashing methods rely on expensive, task-specific training, hindering rapid and scalable retrieval. This paper proposes a training-free multimodal hashing strong baseline: it freezes pre-trained vision and audio encoders to extract features, applies principal component analysis (PCA) for dimensionality reduction, enhances discriminability via random orthogonal projection, and generates compact binary hash codes through threshold-based binarization. Crucially, the method entirely avoids fine-tuning, achieving both computational efficiency and broad cross-domain applicability. Evaluated on standard image retrieval benchmarks and a newly constructed audio hashing benchmark, it attains competitive performance—demonstrating that classical unsupervised techniques retain significant potential in the era of large-scale pre-training. Our approach establishes a novel paradigm for lightweight, plug-and-play cross-modal hashing.

Technology Category

• Machine Learning: Multimodal Learning
• Computer Vision: Multi-modal Vision
• Data Mining & Knowledge Management: Mining of Visual, Multimedia & Multimodal Data

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

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.

Problem

• Training-free hashing for scalable fast search
• Leveraging pretrained encoders for binary embeddings
• Evaluating method on image and audio retrieval benchmarks

Innovation

• Leveraging pretrained encoders for embeddings
• Combining classical training-free hashing techniques
• Using frozen embeddings without fine-tuning