Object-guided multi-granularity unsupervised hashing for image retrieval.

Abstract

Hashing has attracted great attention for large-scale image retrieval due to its low storage cost and high search efficiency. However, most unsupervised hashing methods construct the affinity matrix according to the cosine similarity, which achieves poor performance to measure the semantic similarities among the images. On the other hand, contrastive learning based unsupervised hashing faces the challenge to leverage the positive and negative examples which may be unbalanced and fallible. In this paper, we propose a novel Object-guided Multi-granularity Unsupervised Hashing (OMUH) method based on objective detection. First, a pre-trained objective detection model is applied to the images to produce their pseudo labels which can assist with the reconstruction of the affinity matrix to highlight the subject role of the objects. Then, we enlarge the positive examples and purify the negative examples based on the pseudo labels for contrastive learning. Finally, we design a multi-granularity objective function in consideration of preserving intra-image, inter-image and intra-class similarities for hash codes learning. Extensive experiments on three common benchmarks demonstrate that the proposed model consistently outperforms state-of-the-art unsupervised hashing methods with up to 12% mAP improvement. Our codes are available at https://github.com/caoyuan618/OMUH.