Unsupervised Hashing Network with Hyper Quantization Tree


Sungeun Kim (Ajou University), Jongbin Ryu (Ajou University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Unsupervised hashing network commonly uses pseudo labels generated from a clustering algorithm. Therefore, the performance of the hashing network is completely oriented from the clustering algorithm, so there is no way to overcome inaccurate clustering results. To address this issue, we introduce a hyper-quantization tree method that regularizes erroneous clustering results for training robust unsupervised hashing networks. The proposed method employs a tree structure that forces leaf nodes to merge into two clusters. We refer to this forced merging method as hyper-quantization, and the merged binary cluster is used as pseudo labels that overcome the erroneous clustering results. With this hyper-quantization, we train each bit of the hash code, allowing each bit to learn a diverse feature space. As a result, our hashing network performs better due to the accurate and diversified feature representation. In our experiments, we demonstrate that the proposed hyper-quantization tree greatly enhances the performance of the state-of-the-art unsupervised hashing networks. We also provide in-depth analyses to support our claims on the diversified feature representation. Our code is publicly available at https://github.com/Lab-LVM/HQT

Citation

@inproceedings{Kim_2024_BMVC,
author    = {Sungeun Kim and Jongbin Ryu},
title     = {Unsupervised Hashing Network with Hyper Quantization Tree},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0482.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection