Hashing-Baseline: Rethinking Hashing in the Age of Pretrained Models (ICASSP 2026)

👩‍💻 Authors

¹ Ilyass Moummad, ^1,2 Kawtar Zaher, ³ Lukas Rauch, ¹ Alexis Joly

¹ INRIA, LIRMM, Université de Montpellier, France
² Institut National de l’Audiovisuel, France
³ University of Kassel, Germany

---

🏗️ Repository Overview

This repository demonstrates a simple, training-free baseline for retrieval of images and audio. It works by combining deep pre-trained features with traditional hashing steps — PCA, random orthogonal projection, and a thresholding operation — to produce compact binary codes.

The resulting Hashing-Baseline produces compact binary codes that:

• Retain semantic similarity from the pretrained features
• Allow fast similarity search

This setup provides a straightforward reference for comparing more complex or learned hashing methods on large datasets.

---

⚙️ Method Overview

The Hashing-Baseline pipeline consists of four main steps:

1. Feature Extraction
   Extract high-dimensional features from images or audio using pre-trained models.

2. Dimensionality Reduction with PCA
   Apply Principal Component Analysis (PCA) to reduce feature dimensionality while retaining key semantic information.

3. Random Orthogonal Projection and Binarization
   Project PCA-reduced features with a random orthogonal matrix and binarize by checking whether values are greater than 0, resulting in compact 16-bit hash codes.

4. Retrieval
   Perform similarity search in Hamming space using the binary codes.

This simple pipeline shows that even standard hashing techniques can preserve semantic similarity when applied on strong pretrained features, without any additional training.

---

🔎 Retrieval Example

The figure below shows two query images from the Flickr25K dataset and their 5 nearest neighbors, retrieved using features from a SimDINO ViT-B/16 model pretrained on ImageNet-1K (100 epochs).

• Original features → continuous feature vectors extracted from the backbone (CLS token).
• Hashed 16-bit codes → binary codes produced by applying our Hashing-Baseline to the original features.

This demonstrates that Hashing-Baseline preserves semantic similarity, even with compact 16-bit codes.

📂 Repository Structure

• image/ — Contains code for image retrieval hashing.
  Supports datasets: cifar10, flickr25k, and coco.
  Supported models: dfn and DINOv2/SimDINOv2 via their checkpoints (.pth).
  See image/README.md for detailed instructions.

• audio/ — Contains code for audio retrieval hashing.
  Supports datasets: esc50, gtzan, and speechcommands.
  Supports models: clap, ced, and dasheng.
  See audio/README.md for detailed instructions.

---

🚀 Getting Started

1. Clone the repository

   git clone https://github.com/ilyassmoummad/hashing-baseline
   cd hashing-baseline
   

2. Create a Python environment

   python -m venv envhashingbaseline
   source envhashingbaseline/bin/activate
   python -m pip install -r requirements.txt
   

3. Choose your modality folder (image or audio).
4. Follow the instructions in the README file of the respective folder to download and set up the datasets.
5. Run the provided scripts with your desired model, dataset, and hashing parameters.

---

📚 References

Image Models

• DFN: https://arxiv.org/abs/2309.17425
• DINOv2: https://arxiv.org/abs/2304.07193
• SimDINOv2: https://arxiv.org/abs/2502.10385

Audio Models

• CLAP: https://arxiv.org/abs/2211.06687
• CED: https://arxiv.org/abs/2308.11957
• Dasheng: https://arxiv.org/abs/2406.06992

---

📝 To cite this work:

@misc{hashingbaseline,
      title={Hashing-Baseline: Rethinking Hashing in the Age of Pretrained Models}, 
      author={Ilyass Moummad and Kawtar Zaher and Lukas Rauch and Alexis Joly},
      year={2025},
      eprint={2509.14427},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.14427}, 
}