Contrastive Quantization with Code Memory for Unsupervised Image Retrieval

The high efficiency in computation and storage makes hashing (including binary quantization) a common strategy large-scale retrieval systems. To alleviate the reliance on expensive annotations, unsupervised deep becomes an important research problem. This paper provides novel solution to quantization, namely Contrastive Quantization with Code Memory (MeCoQ). Different from existing reconstruction-based strategies, we learn descriptors by contrastive learning, which can better capture discriminative visual semantics. Besides, uncover that codeword diversity regularization is critical prevent learning-based quantization model degeneration. Moreover, introduce code memory module boosts learning lower feature drift than conventional memories. Extensive experiments benchmark datasets show MeCoQ outperforms state-of-the-art methods. configurations are publicly released.