Discretization Is Not Always Better: Rethinking Deep Quantization for Asymmetric Image Retrieval

This work studies asymmetric image retrieval, where lightweight query-side models interact with stronger gallery-side models. It shows that strict binarization can over-constrain the smaller query model, then introduces Deep Correlation Alignment Hashing (DCAH) to align cross-model similarity structure while performing quantization more implicitly. Experiments on multiple benchmark datasets show that the proposed method improves asymmetric retrieval performance and can be integrated as a plug-and-play quantization module.

Recommended citation: Xinze Liu, Dayan Wu, Hengjie Zhu, Chenming Wu, and Pengwen Dai. 2026. Discretization Is Not Always Better: Rethinking Deep Quantization for Asymmetric Image Retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 40(9), 7323-7331. https://doi.org/10.1609/aaai.v40i9.37670 https://doi.org/10.1609/aaai.v40i9.37670