Hashing-Baseline: Rethinking Hashing in the Age of Pretrained Models
This work addresses the scalability and cost issues in hashing for fast search applications, though it is incremental as it combines existing techniques with modern pretrained models.
The paper tackles the problem of expensive, scenario-specific training required by state-of-the-art hashing methods for information retrieval by introducing Hashing-Baseline, a training-free method that leverages pretrained encoders and classical techniques like PCA and random orthogonal projection. The result is competitive retrieval performance on image and audio benchmarks without additional learning or fine-tuning.
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.