Contrastive Self-Supervised Learning as a Strong Baseline for Unsupervised Hashing

Contrastive self-supervised learning has shown to learn representations transferable to a variety of downstream applications, e.g., object detection and classification. While utilizing a contrastive self-supervised objective to learn generalizable features has been much explored, employing it to directly learn binary representations for image search is yet to be studied. This paper presents Contrastive Self-supervised deep Hashing (CSHash), a simple yet effective unsupervised hashing framework aimed at producing compact binary hash codes that preserve semantic similarity between data without relying on human annotations. CSHash is trained on the basis of a contrastive learning objective, pulling together the augmentations of the same sample and keep apart those of different samples in the hash code space. Evaluation on three datasets shows that simply trained by a contrastive self-supervised loss, CSHash is a strong baseline for unsupervised hashing. It yields discriminative, high-quality binary codes and performs comparably to other unsupervised hashing methods. Additionally, we perform thorough analyses on the main components of CSHash to provide a better insight into the framework.