Self-supervised Product Quantization for Deep Unsupervised Image\n Retrieval

Résumé

Supervised deep learning-based hash and vector quantization are enabling fast\nand large-scale image retrieval systems. By fully exploiting label annotations,\nthey are achieving outstanding retrieval performances compared to the\nconventional methods. However, it is painstaking to assign labels precisely for\na vast amount of training data, and also, the annotation process is\nerror-prone. To tackle these issues, we propose the first deep unsupervised\nimage retrieval method dubbed Self-supervised Product Quantization (SPQ)\nnetwork, which is label-free and trained in a self-supervised manner. We design\na Cross Quantized Contrastive learning strategy that jointly learns codewords\nand deep visual descriptors by comparing individually transformed images\n(views). Our method analyzes the image contents to extract descriptive\nfeatures, allowing us to understand image representations for accurate\nretrieval. By conducting extensive experiments on benchmarks, we demonstrate\nthat the proposed method yields state-of-the-art results even without\nsupervised pretraining.\n

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0