Cross-View Code Alignment for Image Hashing in the Age of Foundation Models

👩‍💻 Authors

¹ Ilyass Moummad, ^1,2 Kawtar Zaher, ³ Hervé Goëau, ¹ Alexis Joly

¹ INRIA, LIRMM, Université de Montpellier, France
² Institut National de l’Audiovisuel, France
³ CIRAD, UMR AMAP, Montpellier, Occitanie, France

---

Keywords: Image Hashing · Image Retrieval · Cross-View Alignment · Coding-Rate Maximization · Foundation Models

---

⚡ TL;DR

We propose Cross-View Code Alignment (CroVCA), a simple and universal principle for hashing foundation model embeddings using Binary Cross-Entropy and Coding-Rate Maximization. It unifies unsupervised and supervised hashing under the same framework.

---

📝 Abstract

Efficient large-scale retrieval requires representations that are both compact and discriminative. Foundation models provide powerful visual and multimodal embeddings, but nearest neighbor search in these high-dimensional spaces is computationally expensive. Hashing offers an efficient alternative by enabling fast Hamming distance search with binary codes, yet existing approaches often rely on complex pipelines, multi-term objectives, designs specialized for a single learning paradigm, and long training times. We introduce CroVCA (Cross-View Code Alignment), a simple and unified principle for learning binary codes that remain consistent across semantically aligned views. A single binary cross-entropy loss enforces alignment, while coding-rate maximization serves as an anti-collapse regularizer to promote balanced and diverse codes. To implement this, we design HashCoder, a lightweight MLP hashing network with a final batch normalization layer to enforce balanced codes. HashCoder can be used as a probing head on frozen embeddings or to adapt encoders efficiently via LoRA fine-tuning. Across benchmarks, CroVCA achieves state-of-the-art results in just 5 training epochs. At 16 bits, it performs particularly well—for instance, unsupervised hashing on COCO in under 2 minutes and supervised hashing on ImageNet100 in about 3 minutes on a single GPU. These results highlight CroVCA’s efficiency, adaptability, and broad applicability.

---

🚀 Features

• Unified training for unsupervised + supervised hashing
• Lightweight HashCoder with BatchNorm promoting balanced binary codes
• Optional LoRA fine-tuning for efficient adaptation of foundation models
• Fast training with just 5 epochs on a single GPU

---

📂 Data Preparation

Follow the step in hashing-baseline and deephash to download the data.

---

💻 Installation

git clone https://github.com/ilyassmoummad/CroVCA.git

Install required packages (see requirements.txt) to ensure dependencies like open_clip, transformers, torchvision, etc., are available:

pip install -r requirements.txt

---

📊 Example Usage

Unsupervised hashing on CIFAR-10:

python main.py --encoder dinov3 --hashcoder small --dataset cifar10 --data_dir path_to_data --bitdim 16 --epochs 5 --device cuda:0

---

📝 To cite this work:

@misc{crovca,
      title={Image Hashing via Cross-View Code Alignment in the Age of Foundation Models}, 
      author={Ilyass Moummad and Kawtar Zaher and Hervé Goëau and Alexis Joly},
      year={2025},
      eprint={2510.27584},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.27584}, 
}